mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-14 18:17:44 +00:00
Revise layout of group convolution (#675)
* [What] Remove pure conv int8 instance
[Why] We will never use pure int8 conv in AI, use int8 quantization instead
* Change layout
* Share the kernel parameter
* Support more type of NHWGC for group conv
* Revise client example of conv 2d, use NHWGC layout
* Add instance to cmake
* Revise layout of group conv quantization instance
* Revise layout of external api of group conv quantization
* Revise layout of group conv quantization client example
* Fix clang format
* Add comment to describe meaning of each parameter
[ROCm/composable_kernel commit: 3eecbfb6ec]
This commit is contained in:
rocking
@@ -17,22 +17,22 @@ using InDataType = ck::half_t;
|
||||
using WeiDataType = ck::half_t;
|
||||
using OutDataType = ck::half_t;
|
||||
|
||||
using InLayout = ck::tensor_layout::convolution::GNHWC;
|
||||
using InLayout = ck::tensor_layout::convolution::NHWGC;
|
||||
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
|
||||
using OutLayout = ck::tensor_layout::convolution::GNHWK;
|
||||
using OutLayout = ck::tensor_layout::convolution::NHWGK;
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
static constexpr ck::index_t NumDimSpatial = 2;
|
||||
static constexpr ck::index_t G = 32;
|
||||
static constexpr ck::index_t N = 256;
|
||||
static constexpr ck::index_t K = 192;
|
||||
static constexpr ck::index_t C = 192;
|
||||
static constexpr ck::index_t Y = 3;
|
||||
static constexpr ck::index_t X = 3;
|
||||
static constexpr ck::index_t Hi = 28;
|
||||
static constexpr ck::index_t Wi = 28;
|
||||
static constexpr ck::index_t Ho = 28;
|
||||
static constexpr ck::index_t Wo = 28;
|
||||
static constexpr ck::index_t N = 256; // batch size
|
||||
static constexpr ck::index_t K = 64; // output channel
|
||||
static constexpr ck::index_t C = 32; // input channel (per group)
|
||||
static constexpr ck::index_t Y = 3; // filter H
|
||||
static constexpr ck::index_t X = 3; // filter W
|
||||
static constexpr ck::index_t Hi = 28; // input H
|
||||
static constexpr ck::index_t Wi = 28; // input W
|
||||
static constexpr ck::index_t Ho = 28; // output H
|
||||
static constexpr ck::index_t Wo = 28; // output W
|
||||
|
||||
struct SimpleDeviceMem
|
||||
{
|
||||
@@ -52,50 +52,24 @@ struct SimpleDeviceMem
|
||||
|
||||
int main()
|
||||
{
|
||||
std::array<ck::index_t, NumDimSpatial + 3> in_lengths{G, N, Hi, Wi, C};
|
||||
std::array<ck::index_t, NumDimSpatial + 3> in_strides{0, 0, 0, 0, 1};
|
||||
|
||||
std::array<ck::index_t, NumDimSpatial + 3> wei_lengths{G, K, Y, X, C};
|
||||
std::array<ck::index_t, NumDimSpatial + 3> wei_strides{0, 0, 0, 0, 1};
|
||||
|
||||
std::array<ck::index_t, NumDimSpatial + 3> out_lengths{G, N, Ho, Wo, K};
|
||||
std::array<ck::index_t, NumDimSpatial + 3> out_strides{0, 0, 0, 0, 1};
|
||||
|
||||
std::partial_sum(rbegin(in_lengths),
|
||||
std::prev(rend(in_lengths)),
|
||||
std::next(rbegin(in_strides)),
|
||||
std::multiplies<>{});
|
||||
std::partial_sum(rbegin(wei_lengths),
|
||||
std::prev(rend(wei_lengths)),
|
||||
std::next(rbegin(wei_strides)),
|
||||
std::multiplies<>{});
|
||||
std::partial_sum(rbegin(out_lengths),
|
||||
std::prev(rend(out_lengths)),
|
||||
std::next(rbegin(out_strides)),
|
||||
std::multiplies<>{});
|
||||
|
||||
// transpose GNHWC/GKYXC/GNHWK to GNCHW/GKCYX/GNCHW
|
||||
std::rotate(
|
||||
rbegin(in_lengths), std::next(rbegin(in_lengths)), std::next(rbegin(in_lengths), 3));
|
||||
std::rotate(
|
||||
rbegin(in_strides), std::next(rbegin(in_strides)), std::next(rbegin(in_strides), 3));
|
||||
std::rotate(
|
||||
rbegin(wei_lengths), std::next(rbegin(wei_lengths)), std::next(rbegin(wei_lengths), 3));
|
||||
std::rotate(
|
||||
rbegin(wei_strides), std::next(rbegin(wei_strides)), std::next(rbegin(wei_strides), 3));
|
||||
std::rotate(
|
||||
rbegin(out_lengths), std::next(rbegin(out_lengths)), std::next(rbegin(out_lengths), 3));
|
||||
std::rotate(
|
||||
rbegin(out_strides), std::next(rbegin(out_strides)), std::next(rbegin(out_strides), 3));
|
||||
// We have NHWGC/GKYXC/NHWGK (x, weight, y) in memory space
|
||||
// However, CK's API only accept length and stride with order of GNCHW/GKCYX/GNCHW
|
||||
// Hence, we need to adjust the order of stride
|
||||
std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
|
||||
std::array<ck::index_t, 5> in_strides{C, Hi * Wi * G * C, 1, Wi * G * C, G * C};
|
||||
std::array<ck::index_t, 5> wei_lengths{G, K, C, Y, X};
|
||||
std::array<ck::index_t, 5> wei_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
|
||||
std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
|
||||
std::array<ck::index_t, 5> out_strides{C, Ho * Wo * G * C, 1, Wo * G * C, G * C};
|
||||
|
||||
std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1};
|
||||
std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1};
|
||||
std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1};
|
||||
std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1};
|
||||
|
||||
SimpleDeviceMem in(sizeof(InDataType) * G * N * Hi * Wi * C);
|
||||
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
|
||||
SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
|
||||
SimpleDeviceMem out(sizeof(OutDataType) * G * N * Ho * Wo * K);
|
||||
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
|
||||
|
||||
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial,
|
||||
InLayout,
|
||||
@@ -155,9 +129,9 @@ int main()
|
||||
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
|
||||
|
||||
std::size_t flop = std::size_t(2) * G * N * K * C * Ho * Wo * Y * X;
|
||||
std::size_t num_bytes = sizeof(InDataType) * G * N * Hi * Wi * C +
|
||||
std::size_t num_bytes = sizeof(InDataType) * N * Hi * Wi * G * C +
|
||||
sizeof(WeiDataType) * G * K * Y * X * C +
|
||||
sizeof(OutDataType) * G * N * Ho * Wo * K;
|
||||
sizeof(OutDataType) * N * Ho * Wo * G * K;
|
||||
|
||||
float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
|
||||
float gb_per_sec = num_bytes / 1.E6 / avg_time;
|
||||
|
||||
@@ -17,26 +17,26 @@ using BiasDataType = int32_t;
|
||||
using RequantScaleDataType = float;
|
||||
using OutDataType = int8_t;
|
||||
|
||||
using InLayout = ck::tensor_layout::convolution::GNHWC;
|
||||
using InLayout = ck::tensor_layout::convolution::NHWGC;
|
||||
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
|
||||
using BiasLayout = ck::tensor_layout::convolution::G_K;
|
||||
using RequantScaleLayout = ck::tensor_layout::convolution::G_K;
|
||||
using OutLayout = ck::tensor_layout::convolution::GNHWK;
|
||||
using OutLayout = ck::tensor_layout::convolution::NHWGK;
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
using ActivationOp = ck::tensor_operation::element_wise::Relu;
|
||||
using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul2_Clamp<ActivationOp>;
|
||||
|
||||
static constexpr ck::index_t NumDimSpatial = 2;
|
||||
static constexpr ck::index_t G = 1;
|
||||
static constexpr ck::index_t N = 4; // batch size
|
||||
static constexpr ck::index_t K = 64; // output channel
|
||||
static constexpr ck::index_t C = 192; // input channel
|
||||
static constexpr ck::index_t Y = 3; // filter H
|
||||
static constexpr ck::index_t X = 3; // filter W
|
||||
static constexpr ck::index_t Hi = 71; // input H
|
||||
static constexpr ck::index_t Wi = 71; // input W
|
||||
static constexpr ck::index_t Ho = 36; // output H
|
||||
static constexpr ck::index_t Wo = 36; // output W
|
||||
static constexpr ck::index_t G = 4;
|
||||
static constexpr ck::index_t N = 4; // batch size
|
||||
static constexpr ck::index_t K = 32; // output channel
|
||||
static constexpr ck::index_t C = 64; // input channel (per group)
|
||||
static constexpr ck::index_t Y = 3; // filter H
|
||||
static constexpr ck::index_t X = 3; // filter W
|
||||
static constexpr ck::index_t Hi = 71; // input H
|
||||
static constexpr ck::index_t Wi = 71; // input W
|
||||
static constexpr ck::index_t Ho = 36; // output H
|
||||
static constexpr ck::index_t Wo = 36; // output W
|
||||
struct SimpleDeviceMem
|
||||
{
|
||||
SimpleDeviceMem() = delete;
|
||||
@@ -55,8 +55,11 @@ struct SimpleDeviceMem
|
||||
|
||||
int main(int argc, char* argv[])
|
||||
{
|
||||
// We have NHWGC/GKYXC/NHWGK (x, weight, y) in memory space
|
||||
// However, CK's API only accept length and stride with order of GNCHW/GKCYX/GNCHW
|
||||
// Hence, we need to adjust the order of stride
|
||||
std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
|
||||
std::array<ck::index_t, 5> in_strides{N * Hi * Wi * C, Hi * Wi * C, 1, Wi * C, C};
|
||||
std::array<ck::index_t, 5> in_strides{C, Hi * Wi * G * C, 1, Wi * G * C, G * C};
|
||||
std::array<ck::index_t, 5> weight_lengths{G, K, C, Y, X};
|
||||
std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
|
||||
std::array<ck::index_t, 5> bias_lengths{G, N, K, Ho, Wo};
|
||||
@@ -64,17 +67,18 @@ int main(int argc, char* argv[])
|
||||
std::array<ck::index_t, 5> requant_scale_lengths{G, N, K, Ho, Wo};
|
||||
std::array<ck::index_t, 5> requant_scale_strides{K, 0, 1, 0, 0};
|
||||
std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
|
||||
std::array<ck::index_t, 5> out_strides{N * Ho * Wo * K, Ho * Wo * K, 1, Wo * K, K};
|
||||
std::array<ck::index_t, 5> out_strides{C, Ho * Wo * G * C, 1, Wo * G * C, G * C};
|
||||
|
||||
std::array<ck::index_t, 2> in_left_pad{1, 1};
|
||||
std::array<ck::index_t, 2> in_right_pad{1, 1};
|
||||
std::array<ck::index_t, 2> conv_strides{2, 2};
|
||||
std::array<ck::index_t, 2> conv_dilations{1, 1};
|
||||
|
||||
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * C);
|
||||
SimpleDeviceMem wei(sizeof(WeiDataType) * K * Y * X * C);
|
||||
SimpleDeviceMem bias(sizeof(BiasDataType) * K * Y * X * C);
|
||||
SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * K);
|
||||
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * K);
|
||||
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
|
||||
SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
|
||||
SimpleDeviceMem bias(sizeof(BiasDataType) * G * K);
|
||||
SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * G * K);
|
||||
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
|
||||
|
||||
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<
|
||||
NumDimSpatial,
|
||||
|
||||
@@ -16,19 +16,19 @@ using WeiDataType = int8_t;
|
||||
using BiasDataType = int32_t;
|
||||
using OutDataType = int8_t;
|
||||
|
||||
using InLayout = ck::tensor_layout::convolution::GNHWC;
|
||||
using InLayout = ck::tensor_layout::convolution::NHWGC;
|
||||
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
|
||||
using BiasLayout = ck::tensor_layout::convolution::G_K;
|
||||
using OutLayout = ck::tensor_layout::convolution::GNHWK;
|
||||
using OutLayout = ck::tensor_layout::convolution::NHWGK;
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
using ActivationOp = ck::tensor_operation::element_wise::Relu;
|
||||
using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul_Clamp<ActivationOp>;
|
||||
|
||||
static constexpr ck::index_t NumDimSpatial = 2;
|
||||
static constexpr ck::index_t G = 1;
|
||||
static constexpr ck::index_t G = 4;
|
||||
static constexpr ck::index_t N = 4; // batch size
|
||||
static constexpr ck::index_t K = 64; // output channel
|
||||
static constexpr ck::index_t C = 192; // input channel
|
||||
static constexpr ck::index_t K = 32; // output channel
|
||||
static constexpr ck::index_t C = 64; // input channel (per group)
|
||||
static constexpr ck::index_t Y = 3; // filter H
|
||||
static constexpr ck::index_t X = 3; // filter W
|
||||
static constexpr ck::index_t Hi = 71; // input H
|
||||
@@ -55,23 +55,27 @@ struct SimpleDeviceMem
|
||||
|
||||
int main(int argc, char* argv[])
|
||||
{
|
||||
// We have NHWGC/GKYXC/NHWGK (x, weight, y) in memory space
|
||||
// However, CK's API only accept length and stride with order of GNCHW/GKCYX/GNCHW
|
||||
// Hence, we need to adjust the order of stride
|
||||
std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
|
||||
std::array<ck::index_t, 5> in_strides{N * Hi * Wi * C, Hi * Wi * C, 1, Wi * C, C};
|
||||
std::array<ck::index_t, 5> in_strides{C, Hi * Wi * G * C, 1, Wi * G * C, G * C};
|
||||
std::array<ck::index_t, 5> weight_lengths{G, K, C, Y, X};
|
||||
std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
|
||||
std::array<ck::index_t, 5> bias_lengths{G, N, K, Ho, Wo};
|
||||
std::array<ck::index_t, 5> bias_strides{K, 0, 1, 0, 0};
|
||||
std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
|
||||
std::array<ck::index_t, 5> out_strides{N * Ho * Wo * K, Ho * Wo * K, 1, Wo * K, K};
|
||||
std::array<ck::index_t, 5> out_strides{C, Ho * Wo * G * C, 1, Wo * G * C, G * C};
|
||||
|
||||
std::array<ck::index_t, 2> in_left_pad{1, 1};
|
||||
std::array<ck::index_t, 2> in_right_pad{1, 1};
|
||||
std::array<ck::index_t, 2> conv_strides{2, 2};
|
||||
std::array<ck::index_t, 2> conv_dilations{1, 1};
|
||||
|
||||
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * C);
|
||||
SimpleDeviceMem wei(sizeof(WeiDataType) * K * Y * X * C);
|
||||
SimpleDeviceMem bias(sizeof(BiasDataType) * K * Y * X * C);
|
||||
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * K);
|
||||
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
|
||||
SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
|
||||
SimpleDeviceMem bias(sizeof(BiasDataType) * G * K);
|
||||
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
|
||||
|
||||
using DeviceOp =
|
||||
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial,
|
||||
|
||||
@@ -17,21 +17,21 @@ using BiasDataType = int32_t;
|
||||
using RequantScaleDataType = float;
|
||||
using OutDataType = int8_t;
|
||||
|
||||
using InLayout = ck::tensor_layout::convolution::GNHWC;
|
||||
using InLayout = ck::tensor_layout::convolution::NHWGC;
|
||||
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
|
||||
using BiasLayout = ck::tensor_layout::convolution::G_K;
|
||||
using RequantScaleLayout = ck::tensor_layout::convolution::G_K;
|
||||
using OutLayout = ck::tensor_layout::convolution::GNHWK;
|
||||
using OutLayout = ck::tensor_layout::convolution::NHWGK;
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
using ActivationOp = ck::tensor_operation::element_wise::TanH;
|
||||
using OutElementOp =
|
||||
ck::tensor_operation::element_wise::Add_Mul2_Activation_Mul_Clamp<ActivationOp>;
|
||||
|
||||
static constexpr ck::index_t NumDimSpatial = 2;
|
||||
static constexpr ck::index_t G = 1;
|
||||
static constexpr ck::index_t G = 4;
|
||||
static constexpr ck::index_t N = 4; // batch size
|
||||
static constexpr ck::index_t K = 64; // output channel
|
||||
static constexpr ck::index_t C = 192; // input channel
|
||||
static constexpr ck::index_t K = 32; // output channel
|
||||
static constexpr ck::index_t C = 64; // input channel (per group)
|
||||
static constexpr ck::index_t Y = 3; // filter H
|
||||
static constexpr ck::index_t X = 3; // filter W
|
||||
static constexpr ck::index_t Hi = 71; // input H
|
||||
@@ -58,8 +58,11 @@ struct SimpleDeviceMem
|
||||
|
||||
int main(int argc, char* argv[])
|
||||
{
|
||||
// We have NHWGC/GKYXC/NHWGK (x, weight, y) in memory space
|
||||
// However, CK's API only accept length and stride with order of GNCHW/GKCYX/GNCHW
|
||||
// Hence, we need to adjust the order of stride
|
||||
std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
|
||||
std::array<ck::index_t, 5> in_strides{N * Hi * Wi * C, Hi * Wi * C, 1, Wi * C, C};
|
||||
std::array<ck::index_t, 5> in_strides{C, Hi * Wi * G * C, 1, Wi * G * C, G * C};
|
||||
std::array<ck::index_t, 5> weight_lengths{G, K, C, Y, X};
|
||||
std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
|
||||
std::array<ck::index_t, 5> bias_lengths{G, N, K, Ho, Wo};
|
||||
@@ -67,17 +70,18 @@ int main(int argc, char* argv[])
|
||||
std::array<ck::index_t, 5> requant_scale_lengths{G, N, K, Ho, Wo};
|
||||
std::array<ck::index_t, 5> requant_scale_strides{K, 0, 1, 0, 0};
|
||||
std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
|
||||
std::array<ck::index_t, 5> out_strides{N * Ho * Wo * K, Ho * Wo * K, 1, Wo * K, K};
|
||||
std::array<ck::index_t, 5> out_strides{C, Ho * Wo * G * C, 1, Wo * G * C, G * C};
|
||||
|
||||
std::array<ck::index_t, 2> in_left_pad{1, 1};
|
||||
std::array<ck::index_t, 2> in_right_pad{1, 1};
|
||||
std::array<ck::index_t, 2> conv_strides{2, 2};
|
||||
std::array<ck::index_t, 2> conv_dilations{1, 1};
|
||||
|
||||
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * C);
|
||||
SimpleDeviceMem wei(sizeof(WeiDataType) * K * Y * X * C);
|
||||
SimpleDeviceMem bias(sizeof(BiasDataType) * K * Y * X * C);
|
||||
SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * K);
|
||||
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * K);
|
||||
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
|
||||
SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
|
||||
SimpleDeviceMem bias(sizeof(BiasDataType) * G * K);
|
||||
SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * G * K);
|
||||
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
|
||||
|
||||
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<
|
||||
NumDimSpatial,
|
||||
|
||||
@@ -16,19 +16,19 @@ using WeiDataType = int8_t;
|
||||
using BiasDataType = int32_t;
|
||||
using OutDataType = int8_t;
|
||||
|
||||
using InLayout = ck::tensor_layout::convolution::GNHWC;
|
||||
using InLayout = ck::tensor_layout::convolution::NHWGC;
|
||||
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
|
||||
using BiasLayout = ck::tensor_layout::convolution::G_K;
|
||||
using OutLayout = ck::tensor_layout::convolution::GNHWK;
|
||||
using OutLayout = ck::tensor_layout::convolution::NHWGK;
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
using ActivationOp = ck::tensor_operation::element_wise::TanH;
|
||||
using OutElementOp = ck::tensor_operation::element_wise::Add_Mul_Activation_Mul_Clamp<ActivationOp>;
|
||||
|
||||
static constexpr ck::index_t NumDimSpatial = 2;
|
||||
static constexpr ck::index_t G = 1;
|
||||
static constexpr ck::index_t G = 4;
|
||||
static constexpr ck::index_t N = 4; // batch size
|
||||
static constexpr ck::index_t K = 64; // output channel
|
||||
static constexpr ck::index_t C = 192; // input channel
|
||||
static constexpr ck::index_t K = 32; // output channel
|
||||
static constexpr ck::index_t C = 64; // input channel (per group)
|
||||
static constexpr ck::index_t Y = 3; // filter H
|
||||
static constexpr ck::index_t X = 3; // filter W
|
||||
static constexpr ck::index_t Hi = 71; // input H
|
||||
@@ -56,23 +56,27 @@ struct SimpleDeviceMem
|
||||
|
||||
int main(int argc, char* argv[])
|
||||
{
|
||||
// We have NHWGC/GKYXC/NHWGK (x, weight, y) in memory space
|
||||
// However, CK's API only accept length and stride with order of GNCHW/GKCYX/GNCHW
|
||||
// Hence, we need to adjust the order of stride
|
||||
std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
|
||||
std::array<ck::index_t, 5> in_strides{N * Hi * Wi * C, Hi * Wi * C, 1, Wi * C, C};
|
||||
std::array<ck::index_t, 5> in_strides{C, Hi * Wi * G * C, 1, Wi * G * C, G * C};
|
||||
std::array<ck::index_t, 5> weight_lengths{G, K, C, Y, X};
|
||||
std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
|
||||
std::array<ck::index_t, 5> bias_lengths{G, N, K, Ho, Wo};
|
||||
std::array<ck::index_t, 5> bias_strides{K, 0, 1, 0, 0};
|
||||
std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
|
||||
std::array<ck::index_t, 5> out_strides{N * Ho * Wo * K, Ho * Wo * K, 1, Wo * K, K};
|
||||
std::array<ck::index_t, 5> out_strides{C, Ho * Wo * G * C, 1, Wo * G * C, G * C};
|
||||
|
||||
std::array<ck::index_t, 2> in_left_pad{1, 1};
|
||||
std::array<ck::index_t, 2> in_right_pad{1, 1};
|
||||
std::array<ck::index_t, 2> conv_strides{2, 2};
|
||||
std::array<ck::index_t, 2> conv_dilations{1, 1};
|
||||
|
||||
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * C);
|
||||
SimpleDeviceMem wei(sizeof(WeiDataType) * K * Y * X * C);
|
||||
SimpleDeviceMem bias(sizeof(BiasDataType) * K * Y * X * C);
|
||||
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * K);
|
||||
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
|
||||
SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
|
||||
SimpleDeviceMem bias(sizeof(BiasDataType) * G * K);
|
||||
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
|
||||
|
||||
using DeviceOp =
|
||||
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial,
|
||||
|
||||
@@ -16,25 +16,25 @@ using WeiDataType = int8_t;
|
||||
using RequantScaleDataType = float;
|
||||
using OutDataType = int8_t;
|
||||
|
||||
using InLayout = ck::tensor_layout::convolution::GNHWC;
|
||||
using InLayout = ck::tensor_layout::convolution::NHWGC;
|
||||
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
|
||||
using RequantScaleLayout = ck::tensor_layout::convolution::G_K;
|
||||
using OutLayout = ck::tensor_layout::convolution::GNHWK;
|
||||
using OutLayout = ck::tensor_layout::convolution::NHWGK;
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
using ActivationOp = PassThrough;
|
||||
using OutElementOp = ck::tensor_operation::element_wise::Activation_Mul2_Clamp<ActivationOp>;
|
||||
|
||||
static constexpr ck::index_t NumDimSpatial = 2;
|
||||
static constexpr ck::index_t G = 1;
|
||||
static constexpr ck::index_t N = 4; // batch size
|
||||
static constexpr ck::index_t K = 64; // output channel
|
||||
static constexpr ck::index_t C = 192; // input channel
|
||||
static constexpr ck::index_t Y = 3; // filter H
|
||||
static constexpr ck::index_t X = 3; // filter W
|
||||
static constexpr ck::index_t Hi = 71; // input H
|
||||
static constexpr ck::index_t Wi = 71; // input W
|
||||
static constexpr ck::index_t Ho = 36; // output H
|
||||
static constexpr ck::index_t Wo = 36; // output W
|
||||
static constexpr ck::index_t G = 4;
|
||||
static constexpr ck::index_t N = 4; // batch size
|
||||
static constexpr ck::index_t K = 32; // output channel
|
||||
static constexpr ck::index_t C = 64; // input channel (per group)
|
||||
static constexpr ck::index_t Y = 3; // filter H
|
||||
static constexpr ck::index_t X = 3; // filter W
|
||||
static constexpr ck::index_t Hi = 71; // input H
|
||||
static constexpr ck::index_t Wi = 71; // input W
|
||||
static constexpr ck::index_t Ho = 36; // output H
|
||||
static constexpr ck::index_t Wo = 36; // output W
|
||||
|
||||
struct SimpleDeviceMem
|
||||
{
|
||||
@@ -54,23 +54,27 @@ struct SimpleDeviceMem
|
||||
|
||||
int main(int argc, char* argv[])
|
||||
{
|
||||
// We have NHWGC/GKYXC/NHWGK (x, weight, y) in memory space
|
||||
// However, CK's API only accept length and stride with order of GNCHW/GKCYX/GNCHW
|
||||
// Hence, we need to adjust the order of stride
|
||||
std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
|
||||
std::array<ck::index_t, 5> in_strides{N * Hi * Wi * C, Hi * Wi * C, 1, Wi * C, C};
|
||||
std::array<ck::index_t, 5> in_strides{C, Hi * Wi * G * C, 1, Wi * G * C, G * C};
|
||||
std::array<ck::index_t, 5> weight_lengths{G, K, C, Y, X};
|
||||
std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
|
||||
std::array<ck::index_t, 5> requant_scale_lengths{G, N, K, Ho, Wo};
|
||||
std::array<ck::index_t, 5> requant_scale_strides{K, 0, 1, 0, 0};
|
||||
std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
|
||||
std::array<ck::index_t, 5> out_strides{N * Ho * Wo * K, Ho * Wo * K, 1, Wo * K, K};
|
||||
std::array<ck::index_t, 5> out_strides{C, Ho * Wo * G * C, 1, Wo * G * C, G * C};
|
||||
|
||||
std::array<ck::index_t, 2> in_left_pad{1, 1};
|
||||
std::array<ck::index_t, 2> in_right_pad{1, 1};
|
||||
std::array<ck::index_t, 2> conv_strides{2, 2};
|
||||
std::array<ck::index_t, 2> conv_dilations{1, 1};
|
||||
|
||||
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * C);
|
||||
SimpleDeviceMem wei(sizeof(WeiDataType) * K * Y * X * C);
|
||||
SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * K);
|
||||
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * K);
|
||||
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
|
||||
SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
|
||||
SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * G * K);
|
||||
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
|
||||
|
||||
using DeviceOp =
|
||||
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial,
|
||||
|
||||
@@ -15,18 +15,18 @@ using InDataType = int8_t;
|
||||
using WeiDataType = int8_t;
|
||||
using OutDataType = int8_t;
|
||||
|
||||
using InLayout = ck::tensor_layout::convolution::GNHWC;
|
||||
using InLayout = ck::tensor_layout::convolution::NHWGC;
|
||||
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
|
||||
using OutLayout = ck::tensor_layout::convolution::GNHWK;
|
||||
using OutLayout = ck::tensor_layout::convolution::NHWGK;
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
using ActivationOp = PassThrough;
|
||||
using OutElementOp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<ActivationOp>;
|
||||
|
||||
static constexpr ck::index_t NumDimSpatial = 2;
|
||||
static constexpr ck::index_t G = 1;
|
||||
static constexpr ck::index_t G = 4;
|
||||
static constexpr ck::index_t N = 4; // batch size
|
||||
static constexpr ck::index_t K = 64; // output channel
|
||||
static constexpr ck::index_t C = 192; // input channel
|
||||
static constexpr ck::index_t K = 32; // output channel
|
||||
static constexpr ck::index_t C = 64; // input channel (per group)
|
||||
static constexpr ck::index_t Y = 3; // filter H
|
||||
static constexpr ck::index_t X = 3; // filter W
|
||||
static constexpr ck::index_t Hi = 71; // input H
|
||||
@@ -53,20 +53,24 @@ struct SimpleDeviceMem
|
||||
|
||||
int main(int argc, char* argv[])
|
||||
{
|
||||
// We have NHWGC/GKYXC/NHWGK (x, weight, y) in memory space
|
||||
// However, CK's API only accept length and stride with order of GNCHW/GKCYX/GNCHW
|
||||
// Hence, we need to adjust the order of stride
|
||||
std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
|
||||
std::array<ck::index_t, 5> in_strides{N * Hi * Wi * C, Hi * Wi * C, 1, Wi * C, C};
|
||||
std::array<ck::index_t, 5> in_strides{C, Hi * Wi * G * C, 1, Wi * G * C, G * C};
|
||||
std::array<ck::index_t, 5> weight_lengths{G, K, C, Y, X};
|
||||
std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
|
||||
std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
|
||||
std::array<ck::index_t, 5> out_strides{N * Ho * Wo * K, Ho * Wo * K, 1, Wo * K, K};
|
||||
std::array<ck::index_t, 5> out_strides{C, Ho * Wo * G * C, 1, Wo * G * C, G * C};
|
||||
|
||||
std::array<ck::index_t, 2> in_left_pad{1, 1};
|
||||
std::array<ck::index_t, 2> in_right_pad{1, 1};
|
||||
std::array<ck::index_t, 2> conv_strides{2, 2};
|
||||
std::array<ck::index_t, 2> conv_dilations{1, 1};
|
||||
|
||||
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * C);
|
||||
SimpleDeviceMem wei(sizeof(WeiDataType) * K * Y * X * C);
|
||||
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * K);
|
||||
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
|
||||
SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
|
||||
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
|
||||
|
||||
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial,
|
||||
InLayout,
|
||||
|
||||
@@ -190,11 +190,11 @@ int run_conv2d_fwd_bias_perchannel_quantization_example(const OutElementOp& out_
|
||||
const auto in_element_op = InElementOp{};
|
||||
const auto wei_element_op = WeiElementOp{};
|
||||
|
||||
using InLayout = ck::tensor_layout::convolution::GNHWC;
|
||||
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
|
||||
using InLayout = ck::tensor_layout::convolution::NHWGC;
|
||||
using WeiLayout = ck::tensor_layout::convolution::KYXGC;
|
||||
using BiasLayout = ck::tensor_layout::convolution::G_K;
|
||||
using RequantScaleLayout = ck::tensor_layout::convolution::G_K;
|
||||
using OutLayout = ck::tensor_layout::convolution::GNHWK;
|
||||
using OutLayout = ck::tensor_layout::convolution::NHWGK;
|
||||
|
||||
const auto in_g_n_c_wis_desc =
|
||||
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param);
|
||||
|
||||
@@ -178,10 +178,10 @@ int run_conv2d_fwd_bias_perlayer_quantization_example(const OutElementOp& out_el
|
||||
const auto in_element_op = InElementOp{};
|
||||
const auto wei_element_op = WeiElementOp{};
|
||||
|
||||
using InLayout = ck::tensor_layout::convolution::GNHWC;
|
||||
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
|
||||
using InLayout = ck::tensor_layout::convolution::NHWGC;
|
||||
using WeiLayout = ck::tensor_layout::convolution::KYXGC;
|
||||
using BiasLayout = ck::tensor_layout::convolution::G_K;
|
||||
using OutLayout = ck::tensor_layout::convolution::GNHWK;
|
||||
using OutLayout = ck::tensor_layout::convolution::NHWGK;
|
||||
|
||||
const auto in_g_n_c_wis_desc =
|
||||
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param);
|
||||
|
||||
@@ -180,10 +180,10 @@ int run_conv2d_fwd_perchannel_quantization_example(const OutElementOp& out_eleme
|
||||
const auto in_element_op = InElementOp{};
|
||||
const auto wei_element_op = WeiElementOp{};
|
||||
|
||||
using InLayout = ck::tensor_layout::convolution::GNHWC;
|
||||
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
|
||||
using InLayout = ck::tensor_layout::convolution::NHWGC;
|
||||
using WeiLayout = ck::tensor_layout::convolution::KYXGC;
|
||||
using RequantScaleLayout = ck::tensor_layout::convolution::G_K;
|
||||
using OutLayout = ck::tensor_layout::convolution::GNHWK;
|
||||
using OutLayout = ck::tensor_layout::convolution::NHWGK;
|
||||
|
||||
const auto in_g_n_c_wis_desc =
|
||||
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param);
|
||||
|
||||
@@ -162,9 +162,9 @@ int run_conv2d_fwd_perlayer_quantization_example(const OutElementOp& out_element
|
||||
const auto in_element_op = InElementOp{};
|
||||
const auto wei_element_op = WeiElementOp{};
|
||||
|
||||
using InLayout = ck::tensor_layout::convolution::GNHWC;
|
||||
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
|
||||
using OutLayout = ck::tensor_layout::convolution::GNHWK;
|
||||
using InLayout = ck::tensor_layout::convolution::NHWGC;
|
||||
using WeiLayout = ck::tensor_layout::convolution::KYXGC;
|
||||
using OutLayout = ck::tensor_layout::convolution::NHWGK;
|
||||
|
||||
const auto in_g_n_c_wis_desc =
|
||||
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param);
|
||||
|
||||
@@ -117,20 +117,6 @@ void add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instances(
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
|
||||
void add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
|
||||
GNHWC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
GNHWK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
Empty_Tuple,
|
||||
int8_t,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
|
||||
void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
|
||||
GNHWC,
|
||||
@@ -159,20 +145,21 @@ void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instances(
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
|
||||
void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_instances(
|
||||
// grouped conv2d forward, NHWGC/GKYXC/NHWGK
|
||||
void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
GNHWK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
NHWGK,
|
||||
BF16,
|
||||
BF16,
|
||||
Empty_Tuple,
|
||||
int8_t,
|
||||
BF16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
// grouped conv2d forward, NHWGC/GKYXC/NHWGK
|
||||
|
||||
void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
|
||||
NHWGC,
|
||||
@@ -187,6 +174,20 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
|
||||
void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
NHWGK,
|
||||
F32,
|
||||
F32,
|
||||
Empty_Tuple,
|
||||
F32,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
|
||||
// grouped conv3d forward, GNDHWC/GKZYXC/GNDHWK
|
||||
void add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
|
||||
@@ -385,12 +386,6 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
|
||||
{
|
||||
add_device_grouped_conv1d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
|
||||
is_same_v<OutDataType, int8_t>)
|
||||
{
|
||||
add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_int8_instances(op_ptrs);
|
||||
add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_instances(op_ptrs);
|
||||
}
|
||||
}
|
||||
else if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWGC> &&
|
||||
is_same_v<WeiLayout, GKYXC> && is_same_v<OutLayout, NHWGK>)
|
||||
@@ -398,7 +393,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
|
||||
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
|
||||
is_same_v<OutDataType, float>)
|
||||
{
|
||||
// no instance
|
||||
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instances(op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
|
||||
is_same_v<OutDataType, half_t>)
|
||||
@@ -409,12 +404,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
|
||||
is_same_v<WeiDataType, ck::bhalf_t> &&
|
||||
is_same_v<OutDataType, ck::bhalf_t>)
|
||||
{
|
||||
// no instance
|
||||
}
|
||||
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
|
||||
is_same_v<OutDataType, int8_t>)
|
||||
{
|
||||
// no instance
|
||||
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instances(op_ptrs);
|
||||
}
|
||||
}
|
||||
else if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, GNDHWC> &&
|
||||
|
||||
@@ -17,14 +17,14 @@ namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
// grouped conv2d forward, GNHWC/GKYXC/GNHWK
|
||||
// grouped conv2d forward, NHWGC/GKYXC/NHWGK
|
||||
void add_device_conv2d_dl_bias_perchannel_quantization_int8_instances(
|
||||
std::vector<
|
||||
std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
GK_GK_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
I32_F32_Tuple,
|
||||
@@ -36,10 +36,10 @@ void add_device_conv2d_dl_bias_perchannel_quantization_int8_instances(
|
||||
|
||||
void add_device_conv2d_dl_bias_relu_perchannel_quantization_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
GK_GK_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
I32_F32_Tuple,
|
||||
@@ -52,10 +52,10 @@ void add_device_conv2d_dl_bias_relu_perchannel_quantization_int8_instances(
|
||||
void add_device_conv2d_dl_bias_tanh_perchannel_quantization_int8_instances(
|
||||
std::vector<
|
||||
std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
GK_GK_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
I32_F32_Tuple,
|
||||
@@ -68,10 +68,10 @@ void add_device_conv2d_dl_bias_tanh_perchannel_quantization_int8_instances(
|
||||
void add_device_conv2d_xdl_bias_perchannel_quantization_int8_instances(
|
||||
std::vector<
|
||||
std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
GK_GK_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
I32_F32_Tuple,
|
||||
@@ -83,10 +83,10 @@ void add_device_conv2d_xdl_bias_perchannel_quantization_int8_instances(
|
||||
|
||||
void add_device_conv2d_xdl_bias_relu_perchannel_quantization_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
GK_GK_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
I32_F32_Tuple,
|
||||
@@ -99,10 +99,10 @@ void add_device_conv2d_xdl_bias_relu_perchannel_quantization_int8_instances(
|
||||
void add_device_conv2d_xdl_bias_tanh_perchannel_quantization_int8_instances(
|
||||
std::vector<
|
||||
std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
GK_GK_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
I32_F32_Tuple,
|
||||
@@ -154,9 +154,9 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
|
||||
{
|
||||
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
|
||||
|
||||
if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, GNHWC> &&
|
||||
if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWGC> &&
|
||||
is_same_v<WeiLayout, GKYXC> && is_same_v<DsLayout, GK_GK_Tuple> &&
|
||||
is_same_v<OutLayout, GNHWK>)
|
||||
is_same_v<OutLayout, NHWGK>)
|
||||
{
|
||||
if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
|
||||
is_same_v<DsDataType, I32_F32_Tuple> && is_same_v<OutDataType, int8_t>)
|
||||
@@ -220,9 +220,9 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
|
||||
{
|
||||
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
|
||||
|
||||
if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, GNHWC> &&
|
||||
if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWGC> &&
|
||||
is_same_v<WeiLayout, GKYXC> && is_same_v<DsLayout, GK_GK_Tuple> &&
|
||||
is_same_v<OutLayout, GNHWK>)
|
||||
is_same_v<OutLayout, NHWGK>)
|
||||
{
|
||||
if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
|
||||
is_same_v<DsDataType, I32_F32_Tuple> && is_same_v<OutDataType, int8_t>)
|
||||
|
||||
@@ -17,14 +17,14 @@ namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
// grouped conv2d forward, GNHWC/GKYXC/GNHWK
|
||||
// grouped conv2d forward, NHWGC/GKYXC/NHWGK
|
||||
void add_device_conv2d_dl_bias_perlayer_quantization_int8_instances(
|
||||
std::vector<
|
||||
std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
I32_Tuple,
|
||||
@@ -36,10 +36,10 @@ void add_device_conv2d_dl_bias_perlayer_quantization_int8_instances(
|
||||
|
||||
void add_device_conv2d_dl_bias_relu_perlayer_quantization_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
I32_Tuple,
|
||||
@@ -51,10 +51,10 @@ void add_device_conv2d_dl_bias_relu_perlayer_quantization_int8_instances(
|
||||
|
||||
void add_device_conv2d_dl_bias_tanh_perlayer_quantization_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
I32_Tuple,
|
||||
@@ -67,10 +67,10 @@ void add_device_conv2d_dl_bias_tanh_perlayer_quantization_int8_instances(
|
||||
void add_device_conv2d_xdl_bias_perlayer_quantization_int8_instances(
|
||||
std::vector<
|
||||
std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
I32_Tuple,
|
||||
@@ -82,10 +82,10 @@ void add_device_conv2d_xdl_bias_perlayer_quantization_int8_instances(
|
||||
|
||||
void add_device_conv2d_xdl_bias_relu_perlayer_quantization_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
I32_Tuple,
|
||||
@@ -97,10 +97,10 @@ void add_device_conv2d_xdl_bias_relu_perlayer_quantization_int8_instances(
|
||||
|
||||
void add_device_conv2d_xdl_bias_tanh_perlayer_quantization_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
I32_Tuple,
|
||||
@@ -152,9 +152,9 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
|
||||
{
|
||||
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
|
||||
|
||||
if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, GNHWC> &&
|
||||
if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWGC> &&
|
||||
is_same_v<WeiLayout, GKYXC> && is_same_v<DsLayout, GK_Tuple> &&
|
||||
is_same_v<OutLayout, GNHWK>)
|
||||
is_same_v<OutLayout, NHWGK>)
|
||||
{
|
||||
if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
|
||||
is_same_v<DsDataType, I32_Tuple> && is_same_v<OutDataType, int8_t>)
|
||||
@@ -218,9 +218,9 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
|
||||
{
|
||||
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
|
||||
|
||||
if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, GNHWC> &&
|
||||
if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWGC> &&
|
||||
is_same_v<WeiLayout, GKYXC> && is_same_v<DsLayout, GK_Tuple> &&
|
||||
is_same_v<OutLayout, GNHWK>)
|
||||
is_same_v<OutLayout, NHWGK>)
|
||||
{
|
||||
if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
|
||||
is_same_v<DsDataType, I32_Tuple> && is_same_v<OutDataType, int8_t>)
|
||||
|
||||
@@ -17,13 +17,13 @@ namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
// grouped conv2d forward, GNHWC/GKYXC/GNHWK
|
||||
// grouped conv2d forward, NHWGC/GKYXC/NHWGK
|
||||
void add_device_conv2d_dl_perchannel_quantization_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
F32_Tuple,
|
||||
@@ -35,10 +35,10 @@ void add_device_conv2d_dl_perchannel_quantization_int8_instances(
|
||||
|
||||
void add_device_conv2d_dl_relu_perchannel_quantization_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
F32_Tuple,
|
||||
@@ -50,10 +50,10 @@ void add_device_conv2d_dl_relu_perchannel_quantization_int8_instances(
|
||||
|
||||
void add_device_conv2d_xdl_perchannel_quantization_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
F32_Tuple,
|
||||
@@ -65,10 +65,10 @@ void add_device_conv2d_xdl_perchannel_quantization_int8_instances(
|
||||
|
||||
void add_device_conv2d_xdl_relu_perchannel_quantization_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
F32_Tuple,
|
||||
@@ -119,9 +119,9 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
|
||||
{
|
||||
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
|
||||
|
||||
if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, GNHWC> &&
|
||||
if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWGC> &&
|
||||
is_same_v<WeiLayout, GKYXC> && is_same_v<DsLayout, GK_Tuple> &&
|
||||
is_same_v<OutLayout, GNHWK>)
|
||||
is_same_v<OutLayout, NHWGK>)
|
||||
{
|
||||
if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
|
||||
is_same_v<OutDataType, int8_t>)
|
||||
|
||||
@@ -17,13 +17,13 @@ namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
// grouped conv2d forward, GNHWC/GKYXC/GNHWK
|
||||
// grouped conv2d forward, NHWGC/GKYXC/NHWGK
|
||||
void add_device_conv2d_dl_perlayer_quantization_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
Empty_Tuple,
|
||||
@@ -35,10 +35,10 @@ void add_device_conv2d_dl_perlayer_quantization_int8_instances(
|
||||
|
||||
void add_device_conv2d_dl_relu_perlayer_quantization_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
Empty_Tuple,
|
||||
@@ -50,10 +50,10 @@ void add_device_conv2d_dl_relu_perlayer_quantization_int8_instances(
|
||||
|
||||
void add_device_conv2d_xdl_perlayer_quantization_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
Empty_Tuple,
|
||||
@@ -65,10 +65,10 @@ void add_device_conv2d_xdl_perlayer_quantization_int8_instances(
|
||||
|
||||
void add_device_conv2d_xdl_relu_perlayer_quantization_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
Empty_Tuple,
|
||||
@@ -117,8 +117,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
|
||||
{
|
||||
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
|
||||
|
||||
if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, GNHWC> &&
|
||||
is_same_v<WeiLayout, GKYXC> && is_same_v<OutLayout, GNHWK>)
|
||||
if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWGC> &&
|
||||
is_same_v<WeiLayout, GKYXC> && is_same_v<OutLayout, NHWGK>)
|
||||
{
|
||||
if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
|
||||
is_same_v<OutDataType, int8_t>)
|
||||
|
||||
@@ -3,11 +3,11 @@ add_instance_library(device_grouped_conv2d_fwd_instance
|
||||
device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
|
||||
device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
|
||||
device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
|
||||
device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_int8_instance.cpp
|
||||
# NHWGC, GKYXC, NHWGK
|
||||
device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
|
||||
device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
|
||||
#dl
|
||||
device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
|
||||
#dl
|
||||
device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
|
||||
device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
|
||||
device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_instance.cpp
|
||||
)
|
||||
|
||||
@@ -0,0 +1,53 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
using BF16 = ck::bhalf_t;
|
||||
using F16 = ck::half_t;
|
||||
using F32 = float;
|
||||
|
||||
using Empty_Tuple = ck::Tuple<>;
|
||||
|
||||
template <ck::index_t... Is>
|
||||
using S = ck::Sequence<Is...>;
|
||||
|
||||
using NHWGC = ck::tensor_layout::convolution::NHWGC;
|
||||
using GNHWC = ck::tensor_layout::convolution::GNHWC;
|
||||
|
||||
using GKYXC = ck::tensor_layout::convolution::GKYXC;
|
||||
|
||||
using NHWGK = ck::tensor_layout::convolution::NHWGK;
|
||||
using GNHWK = ck::tensor_layout::convolution::GNHWK;
|
||||
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
static constexpr auto ConvFwdDefault =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
|
||||
|
||||
static constexpr auto ConvFwd1x1P0 =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
|
||||
|
||||
static constexpr auto ConvFwd1x1S1P0 =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
|
||||
|
||||
static constexpr auto ConvFwdOddC =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization::OddC;
|
||||
|
||||
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -1,100 +1,54 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
#include "device_grouped_conv2d_fwd_dl_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
using InDataType = ck::half_t;
|
||||
using WeiDataType = ck::half_t;
|
||||
using AccDataType = float;
|
||||
using OutDataType = ck::half_t;
|
||||
|
||||
using Empty_Tuple = ck::Tuple<>;
|
||||
template <ck::index_t... Is>
|
||||
using S = ck::Sequence<Is...>;
|
||||
|
||||
using InElementOp = ck::tensor_operation::element_wise::PassThrough;
|
||||
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
|
||||
using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
using InLayout = ck::tensor_layout::convolution::GNHWC;
|
||||
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
|
||||
using OutLayout = ck::tensor_layout::convolution::GNHWK;
|
||||
|
||||
static constexpr auto ConvSpec =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
|
||||
static constexpr auto Filter1x1Pad0 =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
|
||||
static constexpr auto Filter1x1Stride1Pad0 =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
|
||||
|
||||
static constexpr auto GemmPadingSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
|
||||
|
||||
using device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instances = std::tuple<
|
||||
// clang-format off
|
||||
// ########################################| NDim| InData| WeiData| MultpleD| OutData| AccData| InLayout| WeiLayout| MultipleD| OutLayout| In| Wei| Out| Convolution| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
|
||||
// ########################################| Spatial| Type| Type| Type| Type| Type| | | Layout| | Elementwise| Elementwise| Elementwise| Forward| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
|
||||
// ########################################| | | | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | |
|
||||
// ########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK< 2, InDataType, WeiDataType, Empty_Tuple, OutDataType, AccDataType, InLayout, WeiLayout, Empty_Tuple, OutLayout, InElementOp, WeiElementOp, OutElementOp, ConvSpec, GemmPadingSpec, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>
|
||||
// clang-format on
|
||||
>;
|
||||
using device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_Filter1x1Pad0_instances = std::tuple<
|
||||
// clang-format off
|
||||
// ########################################| NDim| InData| WeiData| MultpleD| OutData| AccData| InLayout| WeiLayout| MultipleD| OutLayout| In| Wei| Out| Convolution| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
|
||||
// ########################################| Spatial| Type| Type| Type| Type| Type| | | Layout| | Elementwise| Elementwise| Elementwise| Forward| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
|
||||
// ########################################| | | | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | |
|
||||
// ########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK< 2, InDataType, WeiDataType, Empty_Tuple, OutDataType, AccDataType, InLayout, WeiLayout, Empty_Tuple, OutLayout, InElementOp, WeiElementOp, OutElementOp, Filter1x1Pad0, GemmPadingSpec, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
using device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_Filter1x1Stride1Pad0_instances =
|
||||
std::tuple<
|
||||
// clang-format off
|
||||
// ########################################| NDim| InData| WeiData| MultpleD| OutData| AccData| InLayout| WeiLayout| MultipleD| OutLayout| In| Wei| Out| Convolution| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
|
||||
// ########################################| Spatial| Type| Type| Type| Type| Type| | | Layout| | Elementwise| Elementwise| Elementwise| Forward| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
|
||||
// ########################################| | | | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | |
|
||||
// ########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK< 2, InDataType, WeiDataType, Empty_Tuple, OutDataType, AccDataType, InLayout, WeiLayout, Empty_Tuple, OutLayout, InElementOp, WeiElementOp, OutElementOp, Filter1x1Stride1Pad0, GemmPadingSpec, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
|
||||
InLayout,
|
||||
WeiLayout,
|
||||
GNHWC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
OutLayout,
|
||||
InDataType,
|
||||
WeiDataType,
|
||||
GNHWK,
|
||||
F16,
|
||||
F16,
|
||||
Empty_Tuple,
|
||||
OutDataType,
|
||||
InElementOp,
|
||||
WeiElementOp,
|
||||
OutElementOp>>>& instances)
|
||||
F16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instances{});
|
||||
device_grouped_conv2d_fwd_dl_f16_instances<GNHWC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
GNHWK,
|
||||
Empty_Tuple,
|
||||
PassThrough,
|
||||
ConvFwdDefault>{});
|
||||
|
||||
add_device_operation_instances(
|
||||
instances, device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_Filter1x1Pad0_instances{});
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_fwd_dl_f16_instances<GNHWC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
GNHWK,
|
||||
Empty_Tuple,
|
||||
PassThrough,
|
||||
ConvFwd1x1P0>{});
|
||||
|
||||
add_device_operation_instances(
|
||||
instances,
|
||||
device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_Filter1x1Stride1Pad0_instances{});
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_fwd_dl_f16_instances<GNHWC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
GNHWK,
|
||||
Empty_Tuple,
|
||||
PassThrough,
|
||||
ConvFwd1x1S1P0>{});
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
|
||||
@@ -1,104 +1,54 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
#include "device_grouped_conv2d_fwd_dl_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
using InDataType = float;
|
||||
using WeiDataType = float;
|
||||
using AccDataType = float;
|
||||
using OutDataType = float;
|
||||
|
||||
using Empty_Tuple = ck::Tuple<>;
|
||||
template <ck::index_t... Is>
|
||||
using S = ck::Sequence<Is...>;
|
||||
|
||||
using InElementOp = ck::tensor_operation::element_wise::PassThrough;
|
||||
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
|
||||
using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
using InLayout = ck::tensor_layout::convolution::GNHWC;
|
||||
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
|
||||
using OutLayout = ck::tensor_layout::convolution::GNHWK;
|
||||
|
||||
static constexpr auto ConvSpec =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
|
||||
static constexpr auto Filter1x1Pad0 =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
|
||||
static constexpr auto Filter1x1Stride1Pad0 =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
|
||||
|
||||
static constexpr auto GemmPadingSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
|
||||
|
||||
using device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instances = std::tuple<
|
||||
// clang-format off
|
||||
// clang-format off
|
||||
// ########################################| NDim| InData| WeiData| MultpleD| OutData| AccData| InLayout| WeiLayout| MultipleD| OutLayout| In| Wei| Out| Convolution| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
|
||||
// ########################################| Spatial| Type| Type| Type| Type| Type| | | Layout| | Elementwise| Elementwise| Elementwise| Forward| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
|
||||
// ########################################| | | | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | |
|
||||
// ########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK< 2, InDataType, WeiDataType, Empty_Tuple, OutDataType, AccDataType, InLayout, WeiLayout, Empty_Tuple, OutLayout, InElementOp, WeiElementOp, OutElementOp, ConvSpec, GemmPadingSpec, 256, 128, 128, 16, 1, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 1>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<8, 1, 1, 1>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
using device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_Filter1x1Pad0_instances = std::tuple<
|
||||
// clang-format off
|
||||
// clang-format off
|
||||
// ########################################| NDim| InData| WeiData| MultpleD| OutData| AccData| InLayout| WeiLayout| MultipleD| OutLayout| In| Wei| Out| Convolution| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
|
||||
// ########################################| Spatial| Type| Type| Type| Type| Type| | | Layout| | Elementwise| Elementwise| Elementwise| Forward| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
|
||||
// ########################################| | | | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | |
|
||||
// ########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK< 2, InDataType, WeiDataType, Empty_Tuple, OutDataType, AccDataType, InLayout, WeiLayout, Empty_Tuple, OutLayout, InElementOp, WeiElementOp, OutElementOp, Filter1x1Pad0, GemmPadingSpec, 256, 128, 128, 16, 1, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 1>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<8, 1, 1, 1>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
using device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_Filter1x1Stride1Pad0_instances =
|
||||
std::tuple<
|
||||
// clang-format off
|
||||
// clang-format off
|
||||
// ########################################| NDim| InData| WeiData| MultpleD| OutData| AccData| InLayout| WeiLayout| MultipleD| OutLayout| In| Wei| Out| Convolution| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
|
||||
// ########################################| Spatial| Type| Type| Type| Type| Type| | | Layout| | Elementwise| Elementwise| Elementwise| Forward| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
|
||||
// ########################################| | | | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | |
|
||||
// ########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK< 2, InDataType, WeiDataType, Empty_Tuple, OutDataType, AccDataType, InLayout, WeiLayout, Empty_Tuple, OutLayout, InElementOp, WeiElementOp, OutElementOp, Filter1x1Stride1Pad0, GemmPadingSpec, 256, 128, 128, 16, 1, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 1>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<8, 1, 1, 1>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
|
||||
InLayout,
|
||||
WeiLayout,
|
||||
GNHWC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
OutLayout,
|
||||
InDataType,
|
||||
WeiDataType,
|
||||
GNHWK,
|
||||
F32,
|
||||
F32,
|
||||
Empty_Tuple,
|
||||
OutDataType,
|
||||
InElementOp,
|
||||
WeiElementOp,
|
||||
OutElementOp>>>& instances)
|
||||
F32,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instances{});
|
||||
device_grouped_conv2d_fwd_dl_f32_instances<GNHWC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
GNHWK,
|
||||
Empty_Tuple,
|
||||
PassThrough,
|
||||
ConvFwdDefault>{});
|
||||
|
||||
add_device_operation_instances(
|
||||
instances, device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_Filter1x1Pad0_instances{});
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_fwd_dl_f32_instances<GNHWC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
GNHWK,
|
||||
Empty_Tuple,
|
||||
PassThrough,
|
||||
ConvFwd1x1P0>{});
|
||||
|
||||
add_device_operation_instances(
|
||||
instances,
|
||||
device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_Filter1x1Stride1Pad0_instances{});
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_fwd_dl_f32_instances<GNHWC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
GNHWK,
|
||||
Empty_Tuple,
|
||||
PassThrough,
|
||||
ConvFwd1x1S1P0>{});
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
|
||||
@@ -1,104 +0,0 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
using InDataType = int8_t;
|
||||
using WeiDataType = int8_t;
|
||||
using AccDataType = int32_t;
|
||||
using OutDataType = int8_t;
|
||||
|
||||
using Empty_Tuple = ck::Tuple<>;
|
||||
template <ck::index_t... Is>
|
||||
using S = ck::Sequence<Is...>;
|
||||
|
||||
using InElementOp = ck::tensor_operation::element_wise::PassThrough;
|
||||
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
|
||||
using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
using InLayout = ck::tensor_layout::convolution::GNHWC;
|
||||
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
|
||||
using OutLayout = ck::tensor_layout::convolution::GNHWK;
|
||||
|
||||
static constexpr auto ConvSpec =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
|
||||
static constexpr auto Filter1x1Pad0 =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
|
||||
static constexpr auto Filter1x1Stride1Pad0 =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
|
||||
|
||||
static constexpr auto GemmPadingSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
|
||||
|
||||
using device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_instances = std::tuple<
|
||||
// clang-format off
|
||||
// ########################################| NDim| InData| WeiData| MultpleD| OutData| AccData| InLayout| WeiLayout| MultipleD| OutLayout| In| Wei| Out| Convolution| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
|
||||
// ########################################| Spatial| Type| Type| Type| Type| Type| | | Layout| | Elementwise| Elementwise| Elementwise| Forward| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
|
||||
// ########################################| | | | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | |
|
||||
// ########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK< 2, InDataType, WeiDataType, Empty_Tuple, OutDataType, AccDataType, InLayout, WeiLayout, Empty_Tuple, OutLayout, InElementOp, WeiElementOp, OutElementOp, ConvSpec, GemmPadingSpec, 256, 128, 128, 16, 4, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
using device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_Filter1x1Pad0_instances = std::tuple<
|
||||
// clang-format off
|
||||
// ########################################| NDim| InData| WeiData| MultpleD| OutData| AccData| InLayout| WeiLayout| MultipleD| OutLayout| In| Wei| Out| Convolution| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
|
||||
// ########################################| Spatial| Type| Type| Type| Type| Type| | | Layout| | Elementwise| Elementwise| Elementwise| Forward| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
|
||||
// ########################################| | | | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | |
|
||||
// ########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK< 2, InDataType, WeiDataType, Empty_Tuple, OutDataType, AccDataType, InLayout, WeiLayout, Empty_Tuple, OutLayout, InElementOp, WeiElementOp, OutElementOp, Filter1x1Pad0, GemmPadingSpec, 256, 128, 128, 16, 4, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
using device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_Filter1x1Stride1Pad0_instances =
|
||||
std::tuple<
|
||||
// clang-format off
|
||||
// ########################################| NDim| InData| WeiData| MultpleD| OutData| AccData| InLayout| WeiLayout| MultipleD| OutLayout| In| Wei| Out| Convolution| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
|
||||
// ########################################| Spatial| Type| Type| Type| Type| Type| | | Layout| | Elementwise| Elementwise| Elementwise| Forward| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
|
||||
// ########################################| | | | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | |
|
||||
// ########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK< 2, InDataType, WeiDataType, Empty_Tuple, OutDataType, AccDataType, InLayout, WeiLayout, Empty_Tuple, OutLayout, InElementOp, WeiElementOp, OutElementOp, Filter1x1Stride1Pad0, GemmPadingSpec, 256, 128, 128, 16, 4, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
|
||||
InLayout,
|
||||
WeiLayout,
|
||||
Empty_Tuple,
|
||||
OutLayout,
|
||||
InDataType,
|
||||
WeiDataType,
|
||||
Empty_Tuple,
|
||||
OutDataType,
|
||||
InElementOp,
|
||||
WeiElementOp,
|
||||
OutElementOp>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_instances{});
|
||||
|
||||
add_device_operation_instances(
|
||||
instances, device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_Filter1x1Pad0_instances{});
|
||||
|
||||
add_device_operation_instances(
|
||||
instances,
|
||||
device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_Filter1x1Stride1Pad0_instances{});
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,50 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
|
||||
#include "device_grouped_conv2d_fwd_common.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
template <typename InLayout,
|
||||
typename WeiLayout,
|
||||
typename DsLayout,
|
||||
typename OutLayout,
|
||||
typename DsDatatype,
|
||||
typename CDEElementOp,
|
||||
ConvolutionForwardSpecialization ConvSpec>
|
||||
using device_grouped_conv2d_fwd_dl_f16_instances = std::tuple<
|
||||
// clang-format off
|
||||
// ########################################| NDim| InData| WeiData| MultpleD| OutData| AccData| InLayout| WeiLayout| MultipleD| OutLayout| In| Wei| Out| Convolution| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
|
||||
// ########################################| Spatial| Type| Type| Type| Type| Type| | | Layout| | Elementwise| Elementwise| Elementwise| Forward| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
|
||||
// ########################################| | | | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | |
|
||||
// ########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK< 2, F16, F16, DsDatatype, F16, F32, InLayout, WeiLayout, DsLayout, OutLayout, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
template <typename InLayout,
|
||||
typename WeiLayout,
|
||||
typename DsLayout,
|
||||
typename OutLayout,
|
||||
typename DsDatatype,
|
||||
typename CDEElementOp,
|
||||
ConvolutionForwardSpecialization ConvSpec>
|
||||
using device_grouped_conv2d_fwd_dl_f32_instances = std::tuple<
|
||||
// clang-format off
|
||||
// clang-format off
|
||||
// ########################################| NDim| InData| WeiData| MultpleD| OutData| AccData| InLayout| WeiLayout| MultipleD| OutLayout| In| Wei| Out| Convolution| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
|
||||
// ########################################| Spatial| Type| Type| Type| Type| Type| | | Layout| | Elementwise| Elementwise| Elementwise| Forward| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
|
||||
// ########################################| | | | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | |
|
||||
// ########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK< 2, F32, F32, DsDatatype, F32, F32, InLayout, WeiLayout, DsLayout, OutLayout, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 256, 128, 128, 16, 1, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 1>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<8, 1, 1, 1>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -1,137 +1,14 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
#include "device_grouped_conv2d_fwd_xdl_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
using BF16 = ck::bhalf_t;
|
||||
using F32 = float;
|
||||
|
||||
using Empty_Tuple = ck::Tuple<>;
|
||||
|
||||
template <ck::index_t... Is>
|
||||
using S = ck::Sequence<Is...>;
|
||||
|
||||
using GNHWC = ck::tensor_layout::convolution::GNHWC;
|
||||
using GKYXC = ck::tensor_layout::convolution::GKYXC;
|
||||
using GNHWK = ck::tensor_layout::convolution::GNHWK;
|
||||
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
static constexpr auto ConvFwdDefault =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
|
||||
|
||||
static constexpr auto ConvFwd1x1P0 =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
|
||||
|
||||
static constexpr auto ConvFwd1x1S1P0 =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
|
||||
|
||||
static constexpr auto ConvFwdOddC =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization::OddC;
|
||||
|
||||
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
|
||||
|
||||
// Compilation parameters for in[g, n, hi, wi, c] * wei[g, k, y, x, c] = out[g, n, ho, wo, k]
|
||||
using device_grouped_conv1d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instances =
|
||||
std::tuple<
|
||||
// clang-format off
|
||||
// Default
|
||||
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
|
||||
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
|
||||
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
|
||||
// Filter1x1Pad0
|
||||
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
|
||||
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
|
||||
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
|
||||
// Filter1x1Stride1Pad0
|
||||
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
|
||||
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
|
||||
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
|
||||
// OddC
|
||||
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
|
||||
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
|
||||
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 4, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 4, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 4, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 4, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 4, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 4, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 2, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 2, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 4, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 4, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 4, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 4, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 2, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 2, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 2, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 2, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<2, 32, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<2, 32, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<2, 32, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<2, 32, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 256, 256, 64, 32, 8, 8, 32, 32, 4, 1, S<2, 32, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<2, 32, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<2, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<2, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, BF16, BF16, F32, BF16, Empty_Tuple, BF16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 128, 64, 64, 32, 8, 8, 32, 32, 1, 2, S<2, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<2, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 4>, 8>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
void add_device_grouped_conv1d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
|
||||
GNHWC,
|
||||
@@ -146,8 +23,41 @@ void add_device_grouped_conv1d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instances(
|
||||
PassThrough,
|
||||
PassThrough>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(
|
||||
instances, device_grouped_conv1d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instances{});
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_fwd_xdl_bf16_instances<GNHWC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
GNHWK,
|
||||
Empty_Tuple,
|
||||
PassThrough,
|
||||
ConvFwdDefault>{});
|
||||
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_fwd_xdl_bf16_instances<GNHWC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
GNHWK,
|
||||
Empty_Tuple,
|
||||
PassThrough,
|
||||
ConvFwd1x1P0>{});
|
||||
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_fwd_xdl_bf16_instances<GNHWC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
GNHWK,
|
||||
Empty_Tuple,
|
||||
PassThrough,
|
||||
ConvFwd1x1S1P0>{});
|
||||
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_fwd_xdl_bf16_instances<GNHWC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
GNHWK,
|
||||
Empty_Tuple,
|
||||
PassThrough,
|
||||
ConvFwdOddC>{});
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
|
||||
@@ -1,137 +1,14 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
#include "device_grouped_conv2d_fwd_xdl_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
using F16 = ck::half_t;
|
||||
using F32 = float;
|
||||
|
||||
using Empty_Tuple = ck::Tuple<>;
|
||||
|
||||
template <ck::index_t... Is>
|
||||
using S = ck::Sequence<Is...>;
|
||||
|
||||
using GNHWC = ck::tensor_layout::convolution::GNHWC;
|
||||
using GKYXC = ck::tensor_layout::convolution::GKYXC;
|
||||
using GNHWK = ck::tensor_layout::convolution::GNHWK;
|
||||
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
static constexpr auto ConvFwdDefault =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
|
||||
|
||||
static constexpr auto ConvFwd1x1P0 =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
|
||||
|
||||
static constexpr auto ConvFwd1x1S1P0 =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
|
||||
|
||||
static constexpr auto ConvFwdOddC =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization::OddC;
|
||||
|
||||
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
|
||||
|
||||
// Compilation parameters for in[g, n, hi ,wi, c] * wei[g, k, y, x, c] = out[g, n, ho, wo, k]
|
||||
using device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instances =
|
||||
std::tuple<
|
||||
// clang-format off
|
||||
// Default
|
||||
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
|
||||
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
|
||||
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
|
||||
// Filter1x1Pad0
|
||||
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
|
||||
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
|
||||
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
|
||||
// Filter1x1Stride1Pad0
|
||||
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
|
||||
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
|
||||
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
|
||||
// OddC
|
||||
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
|
||||
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
|
||||
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 4, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 4, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 4, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 4, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 4, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 4, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 2, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 2, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 4, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 4, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 4, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 4, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 2, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 2, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 2, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 2, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<2, 32, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<2, 32, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<2, 32, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<2, 32, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 256, 256, 64, 32, 8, 8, 32, 32, 4, 1, S<2, 32, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<2, 32, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<2, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<2, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 128, 64, 64, 32, 8, 8, 32, 32, 1, 2, S<2, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<2, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 4>, 8>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
void add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
|
||||
GNHWC,
|
||||
@@ -147,7 +24,40 @@ void add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instances(
|
||||
PassThrough>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instances{});
|
||||
device_grouped_conv2d_fwd_xdl_f16_instances<GNHWC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
GNHWK,
|
||||
Empty_Tuple,
|
||||
PassThrough,
|
||||
ConvFwdDefault>{});
|
||||
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_fwd_xdl_f16_instances<GNHWC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
GNHWK,
|
||||
Empty_Tuple,
|
||||
PassThrough,
|
||||
ConvFwd1x1P0>{});
|
||||
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_fwd_xdl_f16_instances<GNHWC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
GNHWK,
|
||||
Empty_Tuple,
|
||||
PassThrough,
|
||||
ConvFwd1x1S1P0>{});
|
||||
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_fwd_xdl_f16_instances<GNHWC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
GNHWK,
|
||||
Empty_Tuple,
|
||||
PassThrough,
|
||||
ConvFwdOddC>{});
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
|
||||
@@ -1,109 +1,14 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
#include "device_grouped_conv2d_fwd_xdl_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
using F32 = float;
|
||||
|
||||
using Empty_Tuple = ck::Tuple<>;
|
||||
|
||||
template <ck::index_t... Is>
|
||||
using S = ck::Sequence<Is...>;
|
||||
|
||||
using GNHWC = ck::tensor_layout::convolution::GNHWC;
|
||||
using GKYXC = ck::tensor_layout::convolution::GKYXC;
|
||||
using GNHWK = ck::tensor_layout::convolution::GNHWK;
|
||||
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
static constexpr auto ConvFwdDefault =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
|
||||
|
||||
static constexpr auto ConvFwd1x1P0 =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
|
||||
|
||||
static constexpr auto ConvFwd1x1S1P0 =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
|
||||
|
||||
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
|
||||
|
||||
// Compilation parameters for in[g, n, hi, wi, c] * wei[g, k, y, x, c] = out[g, n, ho, wo, k]
|
||||
using device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instances =
|
||||
std::tuple<
|
||||
// clang-format off
|
||||
// Default
|
||||
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
|
||||
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
|
||||
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 256, 16, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 128, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 128, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 64, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 64, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 64, 16, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 64, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 128, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 32, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 64, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 32, 64, 16, 4, 4, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>,
|
||||
|
||||
// Filter1x1Pad0
|
||||
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
|
||||
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
|
||||
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 256, 16, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 128, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 128, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 64, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 64, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 64, 16, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 64, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 128, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 32, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 64, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 32, 64, 16, 4, 4, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>,
|
||||
|
||||
// Filter1x1Stride1Pad0
|
||||
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
|
||||
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
|
||||
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 256, 16, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 128, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 128, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 64, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 64, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 64, 16, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 64, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 128, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 32, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 64, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 32, 64, 16, 4, 4, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
void add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
|
||||
GNHWC,
|
||||
@@ -119,7 +24,40 @@ void add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instances(
|
||||
PassThrough>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instances{});
|
||||
device_grouped_conv2d_fwd_xdl_f32_instances<GNHWC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
GNHWK,
|
||||
Empty_Tuple,
|
||||
PassThrough,
|
||||
ConvFwdDefault>{});
|
||||
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_fwd_xdl_f32_instances<GNHWC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
GNHWK,
|
||||
Empty_Tuple,
|
||||
PassThrough,
|
||||
ConvFwd1x1P0>{});
|
||||
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_fwd_xdl_f32_instances<GNHWC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
GNHWK,
|
||||
Empty_Tuple,
|
||||
PassThrough,
|
||||
ConvFwd1x1S1P0>{});
|
||||
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_fwd_xdl_f32_instances<GNHWC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
GNHWK,
|
||||
Empty_Tuple,
|
||||
PassThrough,
|
||||
ConvFwdOddC>{});
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
|
||||
@@ -1,125 +0,0 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
using Empty_Tuple = ck::Tuple<>;
|
||||
|
||||
template <ck::index_t... Is>
|
||||
using S = ck::Sequence<Is...>;
|
||||
|
||||
using GNHWC = ck::tensor_layout::convolution::GNHWC;
|
||||
using GKYXC = ck::tensor_layout::convolution::GKYXC;
|
||||
using GNHWK = ck::tensor_layout::convolution::GNHWK;
|
||||
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
static constexpr auto ConvFwdDefault =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
|
||||
|
||||
static constexpr auto ConvFwd1x1P0 =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
|
||||
|
||||
static constexpr auto ConvFwd1x1S1P0 =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
|
||||
|
||||
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
|
||||
|
||||
// Compilation parameters for in[g, n, hi, wi, c] * wei[g, k, y, x, c] = out[g, n, ho, wo, k]
|
||||
using device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_int8_instances = std::tuple<
|
||||
// clang-format off
|
||||
// Default
|
||||
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
|
||||
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
|
||||
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
|
||||
// Filter1x1Pad0
|
||||
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
|
||||
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
|
||||
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
|
||||
// Filter1x1Stride1Pad0
|
||||
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
|
||||
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
|
||||
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, Empty_Tuple, GNHWK, int8_t, int8_t, int32_t, int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
void add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
|
||||
GNHWC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
GNHWK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
Empty_Tuple,
|
||||
int8_t,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(
|
||||
instances, device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_int8_instances{});
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,105 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
|
||||
#include "device_grouped_conv2d_fwd_common.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
template <typename ALayout,
|
||||
typename BLayout,
|
||||
typename DsLayout,
|
||||
typename ELayout,
|
||||
typename DsDatatype,
|
||||
typename CDEElementOp,
|
||||
ConvolutionForwardSpecialization ConvSpec>
|
||||
using device_grouped_conv2d_fwd_xdl_f16_instances =
|
||||
std::tuple<
|
||||
// clang-format off
|
||||
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
|
||||
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
|
||||
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDatatype, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDatatype, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDatatype, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDatatype, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDatatype, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDatatype, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDatatype, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDatatype, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDatatype, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDatatype, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDatatype, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDatatype, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDatatype, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
template <typename ALayout,
|
||||
typename BLayout,
|
||||
typename DsLayout,
|
||||
typename ELayout,
|
||||
typename DsDatatype,
|
||||
typename CDEElementOp,
|
||||
ConvolutionForwardSpecialization ConvSpec>
|
||||
using device_grouped_conv2d_fwd_xdl_bf16_instances =
|
||||
std::tuple<
|
||||
// clang-format off
|
||||
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
|
||||
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
|
||||
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDatatype, BF16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDatatype, BF16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDatatype, BF16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDatatype, BF16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDatatype, BF16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDatatype, BF16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDatatype, BF16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDatatype, BF16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDatatype, BF16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDatatype, BF16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDatatype, BF16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDatatype, BF16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDatatype, BF16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
template <typename ALayout,
|
||||
typename BLayout,
|
||||
typename DsLayout,
|
||||
typename ELayout,
|
||||
typename DsDatatype,
|
||||
typename CDEElementOp,
|
||||
ConvolutionForwardSpecialization ConvSpec>
|
||||
using device_grouped_conv2d_fwd_xdl_f32_instances =
|
||||
std::tuple<
|
||||
// clang-format off
|
||||
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
|
||||
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
|
||||
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, DsDatatype, F32, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, DsDatatype, F32, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 256, 128, 256, 16, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, DsDatatype, F32, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 128, 128, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, DsDatatype, F32, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, DsDatatype, F32, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 128, 128, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, DsDatatype, F32, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 128, 64, 128, 16, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, DsDatatype, F32, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, DsDatatype, F32, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 256, 128, 64, 16, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, DsDatatype, F32, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 256, 64, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, DsDatatype, F32, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 128, 128, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, DsDatatype, F32, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 128, 32, 128, 16, 4, 4, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, DsDatatype, F32, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 16, 4, 4, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, DsDatatype, F32, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 1, 64, 32, 64, 16, 4, 4, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,66 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
#include "device_grouped_conv2d_fwd_xdl_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
|
||||
void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
NHWGK,
|
||||
BF16,
|
||||
BF16,
|
||||
Empty_Tuple,
|
||||
BF16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_fwd_xdl_bf16_instances<NHWGC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
NHWGK,
|
||||
Empty_Tuple,
|
||||
PassThrough,
|
||||
ConvFwdDefault>{});
|
||||
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_fwd_xdl_bf16_instances<NHWGC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
NHWGK,
|
||||
Empty_Tuple,
|
||||
PassThrough,
|
||||
ConvFwd1x1P0>{});
|
||||
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_fwd_xdl_bf16_instances<NHWGC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
NHWGK,
|
||||
Empty_Tuple,
|
||||
PassThrough,
|
||||
ConvFwd1x1S1P0>{});
|
||||
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_fwd_xdl_bf16_instances<NHWGC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
NHWGK,
|
||||
Empty_Tuple,
|
||||
PassThrough,
|
||||
ConvFwdOddC>{});
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -1,137 +1,14 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
#include "device_grouped_conv2d_fwd_xdl_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
using F16 = ck::half_t;
|
||||
using F32 = float;
|
||||
|
||||
using Empty_Tuple = ck::Tuple<>;
|
||||
|
||||
template <ck::index_t... Is>
|
||||
using S = ck::Sequence<Is...>;
|
||||
|
||||
using NHWGC = ck::tensor_layout::convolution::NHWGC;
|
||||
using GKYXC = ck::tensor_layout::convolution::GKYXC;
|
||||
using NHWGK = ck::tensor_layout::convolution::NHWGK;
|
||||
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
static constexpr auto ConvFwdDefault =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
|
||||
|
||||
static constexpr auto ConvFwd1x1P0 =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
|
||||
|
||||
static constexpr auto ConvFwd1x1S1P0 =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
|
||||
|
||||
static constexpr auto ConvFwdOddC =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization::OddC;
|
||||
|
||||
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
|
||||
|
||||
// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
|
||||
using device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instances =
|
||||
std::tuple<
|
||||
// clang-format off
|
||||
// Default
|
||||
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
|
||||
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
|
||||
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
|
||||
// Filter1x1Pad0
|
||||
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
|
||||
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
|
||||
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1P0, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
|
||||
// Filter1x1Stride1Pad0
|
||||
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
|
||||
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
|
||||
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
|
||||
// OddC
|
||||
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
|
||||
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
|
||||
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 4, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 4, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 4, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 4, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 4, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 4, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 2, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 2, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 8, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 4, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 4, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 4, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 4, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 2, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 2, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 2, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 2, 8>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<2, 32, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<2, 32, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<2, 32, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<2, 32, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 256, 256, 64, 32, 8, 8, 32, 32, 4, 1, S<2, 32, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<2, 32, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 32, 1, 8>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<2, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<2, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 32, 1, 4>, 8>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, NHWGC, GKYXC, Empty_Tuple, NHWGK, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvFwdOddC, GemmMNKPadding, 1, 128, 64, 64, 32, 8, 8, 32, 32, 1, 2, S<2, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<2, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 4>, 8>
|
||||
// clang-format on
|
||||
>;
|
||||
|
||||
void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
|
||||
NHWGC,
|
||||
@@ -147,7 +24,40 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
|
||||
PassThrough>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instances{});
|
||||
device_grouped_conv2d_fwd_xdl_f16_instances<NHWGC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
NHWGK,
|
||||
Empty_Tuple,
|
||||
PassThrough,
|
||||
ConvFwdDefault>{});
|
||||
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_fwd_xdl_f16_instances<NHWGC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
NHWGK,
|
||||
Empty_Tuple,
|
||||
PassThrough,
|
||||
ConvFwd1x1P0>{});
|
||||
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_fwd_xdl_f16_instances<NHWGC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
NHWGK,
|
||||
Empty_Tuple,
|
||||
PassThrough,
|
||||
ConvFwd1x1S1P0>{});
|
||||
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_fwd_xdl_f16_instances<NHWGC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
NHWGK,
|
||||
Empty_Tuple,
|
||||
PassThrough,
|
||||
ConvFwdOddC>{});
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
|
||||
@@ -0,0 +1,66 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
|
||||
#include "device_grouped_conv2d_fwd_xdl_instance.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
|
||||
void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
NHWGK,
|
||||
F32,
|
||||
F32,
|
||||
Empty_Tuple,
|
||||
F32,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_fwd_xdl_f32_instances<NHWGC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
NHWGK,
|
||||
Empty_Tuple,
|
||||
PassThrough,
|
||||
ConvFwdDefault>{});
|
||||
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_fwd_xdl_f32_instances<NHWGC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
NHWGK,
|
||||
Empty_Tuple,
|
||||
PassThrough,
|
||||
ConvFwd1x1P0>{});
|
||||
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_fwd_xdl_f32_instances<NHWGC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
NHWGK,
|
||||
Empty_Tuple,
|
||||
PassThrough,
|
||||
ConvFwd1x1S1P0>{});
|
||||
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_fwd_xdl_f32_instances<NHWGC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
NHWGK,
|
||||
Empty_Tuple,
|
||||
PassThrough,
|
||||
ConvFwdOddC>{});
|
||||
}
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -19,9 +19,9 @@ using Empty_Tuple = ck::Tuple<>;
|
||||
template <ck::index_t... Is>
|
||||
using S = ck::Sequence<Is...>;
|
||||
|
||||
using GNHWC = ck::tensor_layout::convolution::GNHWC;
|
||||
using NHWGC = ck::tensor_layout::convolution::NHWGC;
|
||||
using GKYXC = ck::tensor_layout::convolution::GKYXC;
|
||||
using GNHWK = ck::tensor_layout::convolution::GNHWK;
|
||||
using NHWGK = ck::tensor_layout::convolution::NHWGK;
|
||||
using GK = ck::tensor_layout::convolution::G_K;
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
using Relu = ck::tensor_operation::element_wise::Relu;
|
||||
|
||||
@@ -9,10 +9,10 @@ namespace device {
|
||||
namespace instance {
|
||||
void add_device_conv2d_dl_bias_perchannel_quantization_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
GK_GK_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
I32_F32_Tuple,
|
||||
@@ -23,19 +23,28 @@ void add_device_conv2d_dl_bias_perchannel_quantization_int8_instances(
|
||||
{
|
||||
// dl
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_dl_int8_instances<GK_GK_Tuple,
|
||||
device_grouped_conv2d_dl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_GK_Tuple,
|
||||
NHWGK,
|
||||
I32_F32_Tuple,
|
||||
Add_Mul2_Clamp,
|
||||
ConvFwdDefault,
|
||||
4>{});
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_dl_int8_instances<GK_GK_Tuple,
|
||||
device_grouped_conv2d_dl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_GK_Tuple,
|
||||
NHWGK,
|
||||
I32_F32_Tuple,
|
||||
Add_Mul2_Clamp,
|
||||
ConvFwd1x1P0,
|
||||
4>{});
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_dl_int8_instances<GK_GK_Tuple,
|
||||
device_grouped_conv2d_dl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_GK_Tuple,
|
||||
NHWGK,
|
||||
I32_F32_Tuple,
|
||||
Add_Mul2_Clamp,
|
||||
ConvFwd1x1S1P0,
|
||||
@@ -44,10 +53,10 @@ void add_device_conv2d_dl_bias_perchannel_quantization_int8_instances(
|
||||
|
||||
void add_device_conv2d_dl_bias_relu_perchannel_quantization_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
GK_GK_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
I32_F32_Tuple,
|
||||
@@ -58,19 +67,28 @@ void add_device_conv2d_dl_bias_relu_perchannel_quantization_int8_instances(
|
||||
{
|
||||
// dl
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_dl_int8_instances<GK_GK_Tuple,
|
||||
device_grouped_conv2d_dl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_GK_Tuple,
|
||||
NHWGK,
|
||||
I32_F32_Tuple,
|
||||
Add_Relu_Mul2_Clamp,
|
||||
ConvFwdDefault,
|
||||
4>{});
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_dl_int8_instances<GK_GK_Tuple,
|
||||
device_grouped_conv2d_dl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_GK_Tuple,
|
||||
NHWGK,
|
||||
I32_F32_Tuple,
|
||||
Add_Relu_Mul2_Clamp,
|
||||
ConvFwd1x1P0,
|
||||
4>{});
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_dl_int8_instances<GK_GK_Tuple,
|
||||
device_grouped_conv2d_dl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_GK_Tuple,
|
||||
NHWGK,
|
||||
I32_F32_Tuple,
|
||||
Add_Relu_Mul2_Clamp,
|
||||
ConvFwd1x1S1P0,
|
||||
@@ -79,10 +97,10 @@ void add_device_conv2d_dl_bias_relu_perchannel_quantization_int8_instances(
|
||||
|
||||
void add_device_conv2d_dl_bias_tanh_perchannel_quantization_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
GK_GK_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
I32_F32_Tuple,
|
||||
@@ -93,19 +111,28 @@ void add_device_conv2d_dl_bias_tanh_perchannel_quantization_int8_instances(
|
||||
{
|
||||
// dl
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_dl_int8_instances<GK_GK_Tuple,
|
||||
device_grouped_conv2d_dl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_GK_Tuple,
|
||||
NHWGK,
|
||||
I32_F32_Tuple,
|
||||
Add_Mul2_TanH_Mul_Clamp,
|
||||
ConvFwdDefault,
|
||||
4>{});
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_dl_int8_instances<GK_GK_Tuple,
|
||||
device_grouped_conv2d_dl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_GK_Tuple,
|
||||
NHWGK,
|
||||
I32_F32_Tuple,
|
||||
Add_Mul2_TanH_Mul_Clamp,
|
||||
ConvFwd1x1P0,
|
||||
4>{});
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_dl_int8_instances<GK_GK_Tuple,
|
||||
device_grouped_conv2d_dl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_GK_Tuple,
|
||||
NHWGK,
|
||||
I32_F32_Tuple,
|
||||
Add_Mul2_TanH_Mul_Clamp,
|
||||
ConvFwd1x1S1P0,
|
||||
|
||||
@@ -9,10 +9,10 @@ namespace device {
|
||||
namespace instance {
|
||||
void add_device_conv2d_dl_bias_perlayer_quantization_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
I32_Tuple,
|
||||
@@ -22,19 +22,28 @@ void add_device_conv2d_dl_bias_perlayer_quantization_int8_instances(
|
||||
Add_Mul_Clamp>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_dl_int8_instances<GK_Tuple,
|
||||
device_grouped_conv2d_dl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
NHWGK,
|
||||
I32_Tuple,
|
||||
Add_Mul_Clamp,
|
||||
ConvFwdDefault,
|
||||
4>{});
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_dl_int8_instances<GK_Tuple,
|
||||
device_grouped_conv2d_dl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
NHWGK,
|
||||
I32_Tuple,
|
||||
Add_Mul_Clamp,
|
||||
ConvFwd1x1P0,
|
||||
4>{});
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_dl_int8_instances<GK_Tuple,
|
||||
device_grouped_conv2d_dl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
NHWGK,
|
||||
I32_Tuple,
|
||||
Add_Mul_Clamp,
|
||||
ConvFwd1x1S1P0,
|
||||
@@ -43,10 +52,10 @@ void add_device_conv2d_dl_bias_perlayer_quantization_int8_instances(
|
||||
|
||||
void add_device_conv2d_dl_bias_relu_perlayer_quantization_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
I32_Tuple,
|
||||
@@ -56,21 +65,30 @@ void add_device_conv2d_dl_bias_relu_perlayer_quantization_int8_instances(
|
||||
Add_Relu_Mul_Clamp>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_dl_int8_instances<GK_Tuple,
|
||||
device_grouped_conv2d_dl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
NHWGK,
|
||||
I32_Tuple,
|
||||
Add_Relu_Mul_Clamp,
|
||||
ConvFwdDefault,
|
||||
4>{});
|
||||
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_dl_int8_instances<GK_Tuple,
|
||||
device_grouped_conv2d_dl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
NHWGK,
|
||||
I32_Tuple,
|
||||
Add_Relu_Mul_Clamp,
|
||||
ConvFwd1x1P0,
|
||||
4>{});
|
||||
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_dl_int8_instances<GK_Tuple,
|
||||
device_grouped_conv2d_dl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
NHWGK,
|
||||
I32_Tuple,
|
||||
Add_Relu_Mul_Clamp,
|
||||
ConvFwd1x1S1P0,
|
||||
@@ -79,10 +97,10 @@ void add_device_conv2d_dl_bias_relu_perlayer_quantization_int8_instances(
|
||||
|
||||
void add_device_conv2d_dl_bias_tanh_perlayer_quantization_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
I32_Tuple,
|
||||
@@ -92,21 +110,30 @@ void add_device_conv2d_dl_bias_tanh_perlayer_quantization_int8_instances(
|
||||
Add_Mul_TanH_Mul_Clamp>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_dl_int8_instances<GK_Tuple,
|
||||
device_grouped_conv2d_dl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
NHWGK,
|
||||
I32_Tuple,
|
||||
Add_Mul_TanH_Mul_Clamp,
|
||||
ConvFwdDefault,
|
||||
4>{});
|
||||
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_dl_int8_instances<GK_Tuple,
|
||||
device_grouped_conv2d_dl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
NHWGK,
|
||||
I32_Tuple,
|
||||
Add_Mul_TanH_Mul_Clamp,
|
||||
ConvFwd1x1P0,
|
||||
4>{});
|
||||
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_dl_int8_instances<GK_Tuple,
|
||||
device_grouped_conv2d_dl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
NHWGK,
|
||||
I32_Tuple,
|
||||
Add_Mul_TanH_Mul_Clamp,
|
||||
ConvFwd1x1S1P0,
|
||||
|
||||
@@ -12,7 +12,10 @@ namespace device {
|
||||
namespace instance {
|
||||
|
||||
// clang-format off
|
||||
template <typename DsLayout,
|
||||
template <typename InLayout,
|
||||
typename WeiLayout,
|
||||
typename DsLayout,
|
||||
typename OutLayout,
|
||||
typename DsDatatype,
|
||||
typename OutElementOp,
|
||||
ConvolutionForwardSpecialization ConvSpec,
|
||||
@@ -23,7 +26,7 @@ using device_grouped_conv2d_dl_int8_instances =
|
||||
// ###########################################| Spatial| Type| Type| Type| Type| Type| | | Layout| | Elementwise| Elementwise| Elementwise| Forward| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
|
||||
// ###########################################| | | | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | |
|
||||
// ###########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK< NDimSpatial, int8_t, int8_t, DsDatatype, int8_t, int32_t, GNHWC, GKYXC, DsLayout, GNHWK, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 256, 128, 128, 16, 4, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, DstScalarPerVector>
|
||||
DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK< NDimSpatial, int8_t, int8_t, DsDatatype, int8_t, int32_t, InLayout, WeiLayout, DsLayout, OutLayout, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 256, 128, 128, 16, 4, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, DstScalarPerVector>
|
||||
>;
|
||||
// clang-format on
|
||||
|
||||
|
||||
@@ -9,10 +9,10 @@ namespace device {
|
||||
namespace instance {
|
||||
void add_device_conv2d_dl_perchannel_quantization_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
F32_Tuple,
|
||||
@@ -22,19 +22,28 @@ void add_device_conv2d_dl_perchannel_quantization_int8_instances(
|
||||
Mul2_Clamp>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_dl_int8_instances<GK_Tuple,
|
||||
device_grouped_conv2d_dl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
NHWGK,
|
||||
F32_Tuple,
|
||||
Mul2_Clamp,
|
||||
ConvFwdDefault,
|
||||
4>{});
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_dl_int8_instances<GK_Tuple,
|
||||
device_grouped_conv2d_dl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
NHWGK,
|
||||
F32_Tuple,
|
||||
Mul2_Clamp,
|
||||
ConvFwd1x1P0,
|
||||
4>{});
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_dl_int8_instances<GK_Tuple,
|
||||
device_grouped_conv2d_dl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
NHWGK,
|
||||
F32_Tuple,
|
||||
Mul2_Clamp,
|
||||
ConvFwd1x1S1P0,
|
||||
@@ -43,10 +52,10 @@ void add_device_conv2d_dl_perchannel_quantization_int8_instances(
|
||||
|
||||
void add_device_conv2d_dl_relu_perchannel_quantization_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
F32_Tuple,
|
||||
@@ -56,19 +65,28 @@ void add_device_conv2d_dl_relu_perchannel_quantization_int8_instances(
|
||||
Relu_Mul2_Clamp>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_dl_int8_instances<GK_Tuple,
|
||||
device_grouped_conv2d_dl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
NHWGK,
|
||||
F32_Tuple,
|
||||
Relu_Mul2_Clamp,
|
||||
ConvFwdDefault,
|
||||
4>{});
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_dl_int8_instances<GK_Tuple,
|
||||
device_grouped_conv2d_dl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
NHWGK,
|
||||
F32_Tuple,
|
||||
Relu_Mul2_Clamp,
|
||||
ConvFwd1x1P0,
|
||||
4>{});
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_dl_int8_instances<GK_Tuple,
|
||||
device_grouped_conv2d_dl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
NHWGK,
|
||||
F32_Tuple,
|
||||
Relu_Mul2_Clamp,
|
||||
ConvFwd1x1S1P0,
|
||||
|
||||
@@ -9,10 +9,10 @@ namespace device {
|
||||
namespace instance {
|
||||
void add_device_conv2d_dl_perlayer_quantization_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
Empty_Tuple,
|
||||
@@ -22,19 +22,28 @@ void add_device_conv2d_dl_perlayer_quantization_int8_instances(
|
||||
Mul_Clamp>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_dl_int8_instances<Empty_Tuple,
|
||||
device_grouped_conv2d_dl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
NHWGK,
|
||||
Empty_Tuple,
|
||||
Mul_Clamp,
|
||||
ConvFwdDefault,
|
||||
4>{});
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_dl_int8_instances<Empty_Tuple,
|
||||
device_grouped_conv2d_dl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
NHWGK,
|
||||
Empty_Tuple,
|
||||
Mul_Clamp,
|
||||
ConvFwd1x1P0,
|
||||
4>{});
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_dl_int8_instances<Empty_Tuple,
|
||||
device_grouped_conv2d_dl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
NHWGK,
|
||||
Empty_Tuple,
|
||||
Mul_Clamp,
|
||||
ConvFwd1x1S1P0,
|
||||
@@ -43,10 +52,10 @@ void add_device_conv2d_dl_perlayer_quantization_int8_instances(
|
||||
|
||||
void add_device_conv2d_dl_relu_perlayer_quantization_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
Empty_Tuple,
|
||||
@@ -56,19 +65,28 @@ void add_device_conv2d_dl_relu_perlayer_quantization_int8_instances(
|
||||
Relu_Mul_Clamp>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_dl_int8_instances<Empty_Tuple,
|
||||
device_grouped_conv2d_dl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
NHWGK,
|
||||
Empty_Tuple,
|
||||
Relu_Mul_Clamp,
|
||||
ConvFwdDefault,
|
||||
4>{});
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_dl_int8_instances<Empty_Tuple,
|
||||
device_grouped_conv2d_dl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
NHWGK,
|
||||
Empty_Tuple,
|
||||
Relu_Mul_Clamp,
|
||||
ConvFwd1x1P0,
|
||||
4>{});
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_dl_int8_instances<Empty_Tuple,
|
||||
device_grouped_conv2d_dl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
NHWGK,
|
||||
Empty_Tuple,
|
||||
Relu_Mul_Clamp,
|
||||
ConvFwd1x1S1P0,
|
||||
|
||||
@@ -9,10 +9,10 @@ namespace device {
|
||||
namespace instance {
|
||||
void add_device_conv2d_xdl_bias_perchannel_quantization_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
GK_GK_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
I32_F32_Tuple,
|
||||
@@ -22,19 +22,28 @@ void add_device_conv2d_xdl_bias_perchannel_quantization_int8_instances(
|
||||
Add_Mul2_Clamp>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_xdl_int8_instances<GK_GK_Tuple,
|
||||
device_grouped_conv2d_xdl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_GK_Tuple,
|
||||
NHWGK,
|
||||
I32_F32_Tuple,
|
||||
Add_Mul2_Clamp,
|
||||
ConvFwdDefault,
|
||||
8>{});
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_xdl_int8_instances<GK_GK_Tuple,
|
||||
device_grouped_conv2d_xdl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_GK_Tuple,
|
||||
NHWGK,
|
||||
I32_F32_Tuple,
|
||||
Add_Mul2_Clamp,
|
||||
ConvFwd1x1P0,
|
||||
8>{});
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_xdl_int8_instances<GK_GK_Tuple,
|
||||
device_grouped_conv2d_xdl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_GK_Tuple,
|
||||
NHWGK,
|
||||
I32_F32_Tuple,
|
||||
Add_Mul2_Clamp,
|
||||
ConvFwd1x1S1P0,
|
||||
@@ -43,10 +52,10 @@ void add_device_conv2d_xdl_bias_perchannel_quantization_int8_instances(
|
||||
|
||||
void add_device_conv2d_xdl_bias_relu_perchannel_quantization_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
GK_GK_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
I32_F32_Tuple,
|
||||
@@ -56,19 +65,28 @@ void add_device_conv2d_xdl_bias_relu_perchannel_quantization_int8_instances(
|
||||
Add_Relu_Mul2_Clamp>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_xdl_int8_instances<GK_GK_Tuple,
|
||||
device_grouped_conv2d_xdl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_GK_Tuple,
|
||||
NHWGK,
|
||||
I32_F32_Tuple,
|
||||
Add_Relu_Mul2_Clamp,
|
||||
ConvFwdDefault,
|
||||
8>{});
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_xdl_int8_instances<GK_GK_Tuple,
|
||||
device_grouped_conv2d_xdl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_GK_Tuple,
|
||||
NHWGK,
|
||||
I32_F32_Tuple,
|
||||
Add_Relu_Mul2_Clamp,
|
||||
ConvFwd1x1P0,
|
||||
8>{});
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_xdl_int8_instances<GK_GK_Tuple,
|
||||
device_grouped_conv2d_xdl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_GK_Tuple,
|
||||
NHWGK,
|
||||
I32_F32_Tuple,
|
||||
Add_Relu_Mul2_Clamp,
|
||||
ConvFwd1x1S1P0,
|
||||
@@ -77,10 +95,10 @@ void add_device_conv2d_xdl_bias_relu_perchannel_quantization_int8_instances(
|
||||
|
||||
void add_device_conv2d_xdl_bias_tanh_perchannel_quantization_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
GK_GK_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
I32_F32_Tuple,
|
||||
@@ -90,19 +108,28 @@ void add_device_conv2d_xdl_bias_tanh_perchannel_quantization_int8_instances(
|
||||
Add_Mul2_TanH_Mul_Clamp>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_xdl_int8_instances<GK_GK_Tuple,
|
||||
device_grouped_conv2d_xdl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_GK_Tuple,
|
||||
NHWGK,
|
||||
I32_F32_Tuple,
|
||||
Add_Mul2_TanH_Mul_Clamp,
|
||||
ConvFwdDefault,
|
||||
8>{});
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_xdl_int8_instances<GK_GK_Tuple,
|
||||
device_grouped_conv2d_xdl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_GK_Tuple,
|
||||
NHWGK,
|
||||
I32_F32_Tuple,
|
||||
Add_Mul2_TanH_Mul_Clamp,
|
||||
ConvFwd1x1P0,
|
||||
8>{});
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_xdl_int8_instances<GK_GK_Tuple,
|
||||
device_grouped_conv2d_xdl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_GK_Tuple,
|
||||
NHWGK,
|
||||
I32_F32_Tuple,
|
||||
Add_Mul2_TanH_Mul_Clamp,
|
||||
ConvFwd1x1S1P0,
|
||||
|
||||
@@ -9,10 +9,10 @@ namespace device {
|
||||
namespace instance {
|
||||
void add_device_conv2d_xdl_bias_perlayer_quantization_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
I32_Tuple,
|
||||
@@ -22,19 +22,28 @@ void add_device_conv2d_xdl_bias_perlayer_quantization_int8_instances(
|
||||
Add_Mul_Clamp>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_xdl_int8_instances<GK_Tuple,
|
||||
device_grouped_conv2d_xdl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
NHWGK,
|
||||
I32_Tuple,
|
||||
Add_Mul_Clamp,
|
||||
ConvFwdDefault,
|
||||
8>{});
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_xdl_int8_instances<GK_Tuple,
|
||||
device_grouped_conv2d_xdl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
NHWGK,
|
||||
I32_Tuple,
|
||||
Add_Mul_Clamp,
|
||||
ConvFwd1x1P0,
|
||||
8>{});
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_xdl_int8_instances<GK_Tuple,
|
||||
device_grouped_conv2d_xdl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
NHWGK,
|
||||
I32_Tuple,
|
||||
Add_Mul_Clamp,
|
||||
ConvFwd1x1S1P0,
|
||||
@@ -43,10 +52,10 @@ void add_device_conv2d_xdl_bias_perlayer_quantization_int8_instances(
|
||||
|
||||
void add_device_conv2d_xdl_bias_relu_perlayer_quantization_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
I32_Tuple,
|
||||
@@ -56,21 +65,30 @@ void add_device_conv2d_xdl_bias_relu_perlayer_quantization_int8_instances(
|
||||
Add_Relu_Mul_Clamp>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_xdl_int8_instances<GK_Tuple,
|
||||
device_grouped_conv2d_xdl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
NHWGK,
|
||||
I32_Tuple,
|
||||
Add_Relu_Mul_Clamp,
|
||||
ConvFwdDefault,
|
||||
8>{});
|
||||
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_xdl_int8_instances<GK_Tuple,
|
||||
device_grouped_conv2d_xdl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
NHWGK,
|
||||
I32_Tuple,
|
||||
Add_Relu_Mul_Clamp,
|
||||
ConvFwd1x1P0,
|
||||
8>{});
|
||||
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_xdl_int8_instances<GK_Tuple,
|
||||
device_grouped_conv2d_xdl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
NHWGK,
|
||||
I32_Tuple,
|
||||
Add_Relu_Mul_Clamp,
|
||||
ConvFwd1x1S1P0,
|
||||
@@ -79,10 +97,10 @@ void add_device_conv2d_xdl_bias_relu_perlayer_quantization_int8_instances(
|
||||
|
||||
void add_device_conv2d_xdl_bias_tanh_perlayer_quantization_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
I32_Tuple,
|
||||
@@ -92,21 +110,30 @@ void add_device_conv2d_xdl_bias_tanh_perlayer_quantization_int8_instances(
|
||||
Add_Mul_TanH_Mul_Clamp>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_xdl_int8_instances<GK_Tuple,
|
||||
device_grouped_conv2d_xdl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
NHWGK,
|
||||
I32_Tuple,
|
||||
Add_Mul_TanH_Mul_Clamp,
|
||||
ConvFwdDefault,
|
||||
8>{});
|
||||
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_xdl_int8_instances<GK_Tuple,
|
||||
device_grouped_conv2d_xdl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
NHWGK,
|
||||
I32_Tuple,
|
||||
Add_Mul_TanH_Mul_Clamp,
|
||||
ConvFwd1x1P0,
|
||||
8>{});
|
||||
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_xdl_int8_instances<GK_Tuple,
|
||||
device_grouped_conv2d_xdl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
NHWGK,
|
||||
I32_Tuple,
|
||||
Add_Mul_TanH_Mul_Clamp,
|
||||
ConvFwd1x1S1P0,
|
||||
|
||||
@@ -12,30 +12,33 @@ namespace device {
|
||||
namespace instance {
|
||||
|
||||
// clang-format off
|
||||
template <typename DsLayout,
|
||||
template <typename ALayout,
|
||||
typename BLayout,
|
||||
typename DsLayout,
|
||||
typename ELayout,
|
||||
typename DsDatatype,
|
||||
typename OutElementOp,
|
||||
ConvolutionForwardSpecialization ConvSpec,
|
||||
index_t DstScalarPerVector>
|
||||
using device_grouped_conv2d_xdl_int8_instances =
|
||||
std::tuple <
|
||||
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
|
||||
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
|
||||
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 256, 256, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, DstScalarPerVector>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 256, 128, 256, 64, 16, 16, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, DstScalarPerVector>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 128, 128, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 4>, DstScalarPerVector>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 256, 128, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, DstScalarPerVector>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 128, 128, 64, 64, 16, 16, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 2>, DstScalarPerVector>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 128, 64, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 4>, DstScalarPerVector>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 64, 64, 64, 64, 16, 16, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 2>, DstScalarPerVector>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 256, 128, 64, 64, 16, 16, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, DstScalarPerVector>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 256, 64, 128, 64, 16, 16, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, DstScalarPerVector>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 128, 128, 32, 64, 16, 16, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 2>, DstScalarPerVector>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 128, 32, 128, 64, 16, 16, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 4>, DstScalarPerVector>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 64, 64, 32, 64, 16, 16, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 2>, DstScalarPerVector>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 64, 32, 64, 64, 16, 16, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 2>, DstScalarPerVector>
|
||||
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
|
||||
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
|
||||
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
|
||||
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 256, 256, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, DstScalarPerVector>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 256, 128, 256, 64, 16, 16, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, DstScalarPerVector>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 128, 128, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 4>, DstScalarPerVector>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 256, 128, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, DstScalarPerVector>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 128, 128, 64, 64, 16, 16, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 2>, DstScalarPerVector>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 128, 64, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 4>, DstScalarPerVector>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 64, 64, 64, 64, 16, 16, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 2>, DstScalarPerVector>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 256, 128, 64, 64, 16, 16, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, DstScalarPerVector>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 256, 64, 128, 64, 16, 16, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, DstScalarPerVector>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 128, 128, 32, 64, 16, 16, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 2>, DstScalarPerVector>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 128, 32, 128, 64, 16, 16, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 4>, DstScalarPerVector>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 64, 64, 32, 64, 16, 16, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 2>, DstScalarPerVector>,
|
||||
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 64, 32, 64, 64, 16, 16, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 2>, DstScalarPerVector>
|
||||
>;
|
||||
// clang-format on
|
||||
|
||||
|
||||
@@ -9,10 +9,10 @@ namespace device {
|
||||
namespace instance {
|
||||
void add_device_conv2d_xdl_perchannel_quantization_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
F32_Tuple,
|
||||
@@ -22,19 +22,28 @@ void add_device_conv2d_xdl_perchannel_quantization_int8_instances(
|
||||
Mul2_Clamp>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_xdl_int8_instances<GK_Tuple,
|
||||
device_grouped_conv2d_xdl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
NHWGK,
|
||||
F32_Tuple,
|
||||
Mul2_Clamp,
|
||||
ConvFwdDefault,
|
||||
8>{});
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_xdl_int8_instances<GK_Tuple,
|
||||
device_grouped_conv2d_xdl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
NHWGK,
|
||||
F32_Tuple,
|
||||
Mul2_Clamp,
|
||||
ConvFwd1x1P0,
|
||||
8>{});
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_xdl_int8_instances<GK_Tuple,
|
||||
device_grouped_conv2d_xdl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
NHWGK,
|
||||
F32_Tuple,
|
||||
Mul2_Clamp,
|
||||
ConvFwd1x1S1P0,
|
||||
@@ -43,10 +52,10 @@ void add_device_conv2d_xdl_perchannel_quantization_int8_instances(
|
||||
|
||||
void add_device_conv2d_xdl_relu_perchannel_quantization_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
F32_Tuple,
|
||||
@@ -56,19 +65,28 @@ void add_device_conv2d_xdl_relu_perchannel_quantization_int8_instances(
|
||||
Relu_Mul2_Clamp>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_xdl_int8_instances<GK_Tuple,
|
||||
device_grouped_conv2d_xdl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
NHWGK,
|
||||
F32_Tuple,
|
||||
Relu_Mul2_Clamp,
|
||||
ConvFwdDefault,
|
||||
8>{});
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_xdl_int8_instances<GK_Tuple,
|
||||
device_grouped_conv2d_xdl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
NHWGK,
|
||||
F32_Tuple,
|
||||
Relu_Mul2_Clamp,
|
||||
ConvFwd1x1P0,
|
||||
8>{});
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_xdl_int8_instances<GK_Tuple,
|
||||
device_grouped_conv2d_xdl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
GK_Tuple,
|
||||
NHWGK,
|
||||
F32_Tuple,
|
||||
Relu_Mul2_Clamp,
|
||||
ConvFwd1x1S1P0,
|
||||
|
||||
@@ -9,10 +9,10 @@ namespace device {
|
||||
namespace instance {
|
||||
void add_device_conv2d_xdl_perlayer_quantization_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
Empty_Tuple,
|
||||
@@ -22,19 +22,28 @@ void add_device_conv2d_xdl_perlayer_quantization_int8_instances(
|
||||
Mul_Clamp>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_xdl_int8_instances<Empty_Tuple,
|
||||
device_grouped_conv2d_xdl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
NHWGK,
|
||||
Empty_Tuple,
|
||||
Mul_Clamp,
|
||||
ConvFwdDefault,
|
||||
16>{});
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_xdl_int8_instances<Empty_Tuple,
|
||||
device_grouped_conv2d_xdl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
NHWGK,
|
||||
Empty_Tuple,
|
||||
Mul_Clamp,
|
||||
ConvFwd1x1P0,
|
||||
16>{});
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_xdl_int8_instances<Empty_Tuple,
|
||||
device_grouped_conv2d_xdl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
NHWGK,
|
||||
Empty_Tuple,
|
||||
Mul_Clamp,
|
||||
ConvFwd1x1S1P0,
|
||||
@@ -43,10 +52,10 @@ void add_device_conv2d_xdl_perlayer_quantization_int8_instances(
|
||||
|
||||
void add_device_conv2d_xdl_relu_perlayer_quantization_int8_instances(
|
||||
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
|
||||
GNHWC,
|
||||
NHWGC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
GNHWK,
|
||||
NHWGK,
|
||||
int8_t,
|
||||
int8_t,
|
||||
Empty_Tuple,
|
||||
@@ -56,19 +65,28 @@ void add_device_conv2d_xdl_relu_perlayer_quantization_int8_instances(
|
||||
Relu_Mul_Clamp>>>& instances)
|
||||
{
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_xdl_int8_instances<Empty_Tuple,
|
||||
device_grouped_conv2d_xdl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
NHWGK,
|
||||
Empty_Tuple,
|
||||
Relu_Mul_Clamp,
|
||||
ConvFwdDefault,
|
||||
16>{});
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_xdl_int8_instances<Empty_Tuple,
|
||||
device_grouped_conv2d_xdl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
NHWGK,
|
||||
Empty_Tuple,
|
||||
Relu_Mul_Clamp,
|
||||
ConvFwd1x1P0,
|
||||
16>{});
|
||||
add_device_operation_instances(instances,
|
||||
device_grouped_conv2d_xdl_int8_instances<Empty_Tuple,
|
||||
device_grouped_conv2d_xdl_int8_instances<NHWGC,
|
||||
GKYXC,
|
||||
Empty_Tuple,
|
||||
NHWGK,
|
||||
Empty_Tuple,
|
||||
Relu_Mul_Clamp,
|
||||
ConvFwd1x1S1P0,
|
||||
|
||||
Reference in New Issue
Block a user